Aprendizagem mecânica de previsão de divisas

Jon V.

BigData. Iniciantes. Negociação.

BigData. Iniciantes. Negociação.

Machine Beats Human: Usando Aprendizado de Máquinas em Forex.

A aprendizagem e o comércio de máquinas são um assunto muito interessante. Também é um assunto onde você pode gastar toneladas de tempo escrevendo código e lendo documentos e, em seguida, uma criança pode vencê-lo enquanto joga Mario Kart.

Nas postagens nexts, vamos falar sobre:

Otimize entradas e saídas. Isto e só isso pode fazer uma tonelada de diferença em seu rolo de banco. Calcule o tamanho da posição (no caso de você não gostar do critério de Kelly) Encontre a possível correlação entre diferentes pares (negociação em pares). Adoro a correlação EURUSD vs GBPJPY! Calcule suporte e amp; linhas de resistência.

Mas o que é a Aprendizagem de Máquinas?

Os algoritmos de aprendizagem de máquina são algoritmos em que uma máquina pode identificar padrões em seus dados. Yeap, é assim tão simples. Por exemplo, encontre todos os animais nesta foto e desenhe uma caixa ao redor deles. Além disso, nomeie esse animal. Louco, eu sei. Para negociar como você pode imaginar, é bastante semelhante:

Para que uma máquina "aprenda", você precisa ensinar o que é certo ou errado (aprendizado supervisionado) ou dar-lhe um grande conjunto de dados e deixá-lo ficar selvagem (sem supervisão). Para identificar objetos, isso é direto, e o que é negociar?

Eu olhei em volta para ver se há algum programa de aprendizado de máquina que pode identificar as linhas S / R, mas sem sucesso. Então eu decidi escrever o primeiro programa de aprendizagem de máquinas em python que identifica linhas de suporte e resistência em Python. Outro primeiro! Hooray!

Mas como um algoritmo pode identificar essas áreas? Hoooooow? Senhoras e colegas (e robôs), permitam-me que lhe apresente o MeanShift, um algoritmo sem supervisão que é usado principalmente para o reconhecimento de imagens e que é bastante trivial para configurar e executar (mas também muito lento).

A idéia é que este algoritmo me permita dividir meus dados (txs forex) em áreas e então eu posso usar as "bordas" como suporte e linhas de resistência. Ideia legal, mas isso funciona?

Analisamos cerca de 12 milhões de pontos de dados da EURUSD em 2017 e alguns meses de 2018. As linhas de resistência são colocadas automaticamente por um algoritmo de aprendizado de máquina.

O que é realmente legal (e assustador) é que o algoritmo praticamente o engana. NAILS é difícil. Isso fica realmente assustador quando vamos usar o algoritmo para identificar microestruturas e começar a curar.

O sistema é capaz de processar qualquer tipo de dados temporários (ações, forex, ouro, seja o que for) e renderá um gráfico interativo html (como o gráfico acima) com seus dados e a máquina gerada S / L. O código está aqui, então fique louco.

Agora vamos passar pelo código. Depois de ter seu conjunto de dados, você precisa lê-los e limpá-los. Prepare-se para alguns pandas de magia.

Nós deixamos cair os valores vazios (fins de semana) e depois reescrevemos os dados para candelas de 24 horas (ohcl). Isso torna muito mais fácil traçar. Os dados agrupados são os dados que inseriremos no algoritmo ml.

Em seguida, preparamos os dados que vamos usar no algo.

Na próxima publicação, vamos discutir como melhorar esse trabalho, discutir alguns resultados muito interessantes (o algoritmo pode realmente prever sobre o futuro?) E começar a usá-lo em nossa própria negociação. Se você quiser verificar o próximo artigo e ler mais sobre comércio e investimento usando algoritmos, inscreva-se no boletim informativo.

Próxima próxima: Machine Learning Gone Wild - Usando o código!

Se você tiver mais comentários, clique-me no jonromero ou inscreva-se no boletim informativo.

Legal outro. Este é um tutorial de engenharia sobre como construir uma plataforma algotrading para experimentação e FUN. Qualquer sugestão aqui não é um conselho financeiro. Se você perder qualquer (ou todos) o seu dinheiro porque seguiu quaisquer conselhos de negociação ou implantou este sistema na produção, não pode culpar este blog aleatório (e / ou eu). Aproveite a seu próprio risco.

Construa melhores estratégias! Parte 4: Aprendizado de máquinas.

Deep Blue foi o primeiro computador que ganhou um campeonato mundial de xadrez. Isso foi em 1996 e levou 20 anos até que outro programa, o AlphaGo, pudesse derrotar o melhor jogador Go humano. Deep Blue era um sistema baseado em modelo com regras de xadrez hardwired. O AlphaGo é um sistema de mineração de dados, uma rede neural profunda treinada com milhares de jogos Go. Hardware não melhorado, mas um avanço no software foi essencial para o passo de vencer os melhores jogadores de xadrez para vencer os melhores jogadores Go.

Nesta 4ª parte da mini-série, analisaremos a abordagem de mineração de dados para o desenvolvimento de estratégias comerciais. Este método não se preocupa com os mecanismos de mercado. Ele apenas verifica curvas de preços ou outras fontes de dados para padrões preditivos. Aprendizagem de máquina ou "Inteligência Artificial" e # 8221; nem sempre está envolvido em estratégias de mineração de dados. Na verdade, o mais popular & # 8211; e surpreendentemente lucrativo & # 8211; O método de mineração de dados funciona sem redes neurais sofisticadas ou máquinas de vetor de suporte.

Princípios de aprendizado da máquina.

Um algoritmo de aprendizagem é alimentado com amostras de dados, normalmente derivadas de algum modo de preços históricos. Cada amostra consiste em n variáveis ​​x 1 .. x n, comumente designadas preditores, recursos, sinais ou simplesmente entrada. Esses preditores podem ser os retornos de preços das últimas barras n, ou uma coleção de indicadores clássicos, ou qualquer outra função imaginável da curva de preços (I & # 8217; até mesmo visto os pixels de uma imagem de gráfico de preços usada como preditor para uma neural rede!). Cada amostra também inclui normalmente uma variável alvo y, como o retorno do próximo comércio depois de tirar a amostra, ou o próximo movimento de preços. Na literatura, você pode encontrar também o nome do rótulo ou objetivo. Em um processo de treinamento, o algoritmo aprende a prever o alvo y a partir dos preditores x 1 .. x n. A memória aprendida & # 8216; & # 8217; é armazenado em uma estrutura de dados chamada modelo que é específico para o algoritmo (não deve ser confundido com um modelo financeiro para estratégias baseadas em modelos!). Um modelo de aprendizagem de máquina pode ser uma função com regras de predição no código C, gerado pelo processo de treinamento. Ou pode ser um conjunto de pesos de conexão de uma rede neural.

Os preditores, características, ou o que quer que você os chama, devem conter informações suficientes para prever o alvo e com alguma precisão. Eles também cumprem com freqüência dois requisitos formais. Primeiro, todos os valores de preditores devem estar no mesmo intervalo, como -1 ... +1 (para a maioria dos algoritmos R) ou -100 ... +100 (para algoritmos Zorro ou TSSB). Então você precisa normalizá-los de alguma forma antes de enviá-los para a máquina. Em segundo lugar, as amostras devem ser equilibradas, ou seja, distribuídas igualmente em todos os valores da variável alvo. Então, deve haver quase tantos como ganhar amostras. Se você não observar estes dois requisitos, você se perguntará por que você está obtendo resultados ruins do algoritmo de aprendizado da máquina.

Os algoritmos de regressão prevêem um valor numérico, como a magnitude e o sinal do próximo movimento de preços. Os algoritmos de classificação prevêem uma classe de amostra qualitativa, por exemplo, se ela está precedendo uma vitória ou uma perda. Alguns algoritmos, como redes neurais, árvores de decisão ou máquinas de vetor de suporte, podem ser executados em ambos os modos.

Alguns algoritmos aprendem a dividir amostras em classes sem necessidade de qualquer alvo y. A aprendizagem sem supervisão desse tipo, em oposição à aprendizagem supervisionada usando um alvo. Somewhere inbetween é o aprendizado de reforço, onde o sistema se treina executando simulações com os recursos fornecidos e usando o resultado como alvo de treinamento. AlphaZero, o sucessor do AlphaGo, usou a aprendizagem de reforço ao jogar milhões de jogos Go contra si. Em finanças, há poucas aplicações para aprendizagem sem supervisão ou reforço. 99% das estratégias de aprendizagem de máquinas usam a aprendizagem supervisionada.

Independentemente dos sinais que usamos para preditores em finanças, eles provavelmente contêm muito ruído e pouca informação, e não serão estacionários além disso. Portanto, a previsão financeira é uma das tarefas mais difíceis na aprendizagem por máquinas. Algoritmos mais complexos não conseguem necessariamente melhores resultados. A seleção dos preditores é fundamental para o sucesso. Não é bom usar muitos preditores, uma vez que isso simplesmente causa superação e falha na operação da amostra. Portanto, as estratégias de mineração de dados geralmente aplicam um algoritmo de pré-eleição que determina um pequeno número de preditores de um grupo de muitos. A pré-seleção pode basear-se na correlação entre preditores, na significância, no conteúdo da informação ou simplesmente no sucesso da previsão com um conjunto de testes. Experimentos práticos com seleção de recursos podem ser encontrados em um artigo recente sobre o blog Robot Wealth.

Aqui é uma lista dos métodos de mineração de dados mais populares usados ​​em finanças.

1. Sopa indicadora.

A maioria dos sistemas de negociação que nós estamos programando para clientes não são baseados em um modelo financeiro. O cliente só queria sinais comerciais de certos indicadores técnicos, filtrado com outros indicadores técnicos em combinação com indicadores mais técnicos. Quando perguntado como essa mistura de indicadores poderia ser uma estratégia rentável, ele normalmente respondeu: "Confie em mim". Eu negocie-o manualmente e funciona. & # 8221;

Certamente. Pelo menos às vezes. Embora a maioria desses sistemas não tenha passado um teste WFA (e alguns nem mesmo um backtest simples), um número surpreendentemente grande. E esses também foram geralmente lucrativos no comércio real. O cliente havia experimentado sistematicamente indicadores técnicos até encontrar uma combinação que funcionasse em negociação ao vivo com certos ativos. Esta maneira de análise técnica de teste e erro é uma abordagem clássica de mineração de dados, apenas executada por um ser humano e não por uma máquina. Eu realmente não posso recomendar este método # 8211; e muita sorte, para não falar de dinheiro, provavelmente está envolvido & # 8211; mas posso testemunhar que às vezes leva a sistemas lucrativos.

2. Padrões de velas.

Não deve ser confundido com os padrões japoneses de velas que tiveram a melhor data antes, há muito tempo. O equivalente moderno é a negociação de ações de preço. Você ainda está olhando o aberto, alto, baixo e fechado de velas. Você ainda espera encontrar um padrão que preveja uma direção de preço. Mas você agora está curando curvas de preços contemporâneas para coleta desses padrões. Existem pacotes de software para esse fim. Eles procuram padrões que são lucrativos por algum critério definido pelo usuário, e usá-los para criar uma função de detecção de padrões específica. Poderia parecer este (do analisador de padrão Zorro & # 8217; s):

Esta função C retorna 1 quando os sinais correspondem a um dos padrões, caso contrário, você pode ver do longo código que esta não é a maneira mais rápida de detectar padrões. Um método melhor, usado pelo Zorro quando a função de detecção não precisa ser exportada, é classificar os sinais por sua magnitude e verificar a ordem de classificação. Um exemplo desse sistema pode ser encontrado aqui.

O mercado de ações de preços pode realmente funcionar? Assim como a sopa de indicadores, ela não é baseada em nenhum modelo financeiro racional. Pode-se, na melhor das hipóteses, imaginar que as seqüências de movimentos de preços levem os participantes do mercado a reagirem de uma certa maneira, estabelecendo assim um padrão preditivo temporário. No entanto, o número de padrões é bastante limitado quando você olha apenas as seqüências de algumas velas adjacentes. O próximo passo é comparar velas que não são adjacentes, mas arbitrariamente selecionadas dentro de um período de tempo mais longo. Desta forma, você está obtendo um número quase ilimitado de padrões & # 8211; mas à custa de deixar finalmente o reino do racional. É difícil imaginar como um movimento de preços pode ser previsto por alguns padrões de velas de semanas atrás.

Ainda assim, há muito esforço para isso. Um colega de blogueiro, Daniel Fernandez, administra um site de inscrição (Asirikuy) especializado em padrões de vela de dados minerados. Ele refinou o padrão de negociação até os menores detalhes, e se alguém conseguisse algum lucro desta forma, seria ele. Mas para seus assinantes & # 8217; desapontamento, trocando seus padrões ao vivo (QuriQuant) produziu resultados muito diferentes do que seus maravilhosos backtests. Se os sistemas de ação de preço rentáveis ​​realmente existem, aparentemente ninguém já os encontrou.

3. Regressão linear.

A base simples de muitos algoritmos complexos de aprendizagem de máquina: Prever a variável alvo y por uma combinação linear dos preditores x 1 .. x n.

Os coeficientes a n são o modelo. Eles são calculados para minimizar a soma de diferenças quadradas entre os valores verdadeiros de y das amostras de treino e seus i preditos a partir da fórmula acima:

Para amostras distribuídas normais, a minimização é possível com alguma aritmética da matriz, portanto, nenhuma iteração é necessária. No caso n = 1 & # 8211; com apenas uma variável preditor x & # 8211; a fórmula de regressão é reduzida para.

que é uma regressão linear simples, em oposição à regressão linear multivariada onde n & gt; 1. A regressão linear simples está disponível na maioria das plataformas de negociação, f. i. com o indicador LinReg no TA-Lib. Com y = preço e x = tempo, muitas vezes usado como alternativa para uma média móvel. A regressão linear multivariada está disponível na plataforma R através da função lm (...) que vem com a instalação padrão. Uma variante é a regressão polinomial. Como regressão simples, ele usa apenas uma variável preditor x, mas também seus graus quadrados e superiores, de modo que x n == x n:

Com n = 2 ou n = 3, a regressão polinomial é freqüentemente usada para prever o próximo preço médio a partir dos preços suavizados das últimas barras. A função polyfit de MatLab, R, Zorro e muitas outras plataformas podem ser usadas para regressão polinomial.

4. Perceptron.

Muitas vezes referido como uma rede neural com apenas um neurônio. Na verdade, um perceptron é uma função de regressão como acima, mas com um resultado binário, assim chamado de regressão logística. Não é regressão, é um algoritmo de classificação. A função de recomendação do Zorro (PERCEPTRON, & # 8230;) gera código C que retorna 100 ou -100, dependendo se o resultado previsto está acima de um limite ou não:

Você pode ver que a matriz sig é equivalente às características x n na fórmula de regressão, e os fatores numéricos são os coeficientes a n.

5. Redes nacionais.

A regressão linear ou logística só pode resolver problemas lineares. Muitos não se enquadram nessa categoria & # 8211; um exemplo famoso é prever a saída de uma função XOR simples. E provavelmente também previsão de preços ou retornos comerciais. Uma rede neural artificial (ANN) pode enfrentar problemas não-lineares. É um monte de perceptrons que estão conectados em uma série de camadas. Qualquer perceptron é um neurônio da rede. Sua saída vai para as entradas de todos os neurônios da próxima camada, como esta:

Como o perceptron, uma rede neural também aprende determinando os coeficientes que minimizam o erro entre a previsão da amostra e o alvo da amostra. Mas isso exige agora um processo de aproximação, normalmente com backpropagating o erro da saída para as entradas, otimizando os pesos a caminho. Este processo impõe duas restrições. Primeiro, as saídas do neurônio devem agora ser continuamente funções diferenciáveis ​​em vez do limiar de perceptron simples. Em segundo lugar, a rede não deve ser muito profunda e # 8211; não deve ter muitas camadas escondidas & # 8217; de neurônios entre entradas e saída. Esta segunda restrição limita a complexidade dos problemas que uma rede neural padrão pode resolver.

Ao usar uma rede neural para previsão de negociações, você tem muitos parâmetros com os quais você pode brincar e, se você não for cuidadoso, produza muitos tipos de seleção:

Número de camadas ocultas Número de neurônios por camada oculta Número de ciclos de backpropagation, épocas nomeadas Taxa de aprendizado, a largura do passo de uma Momência de época, um fator de inércia para a função de ativação da adaptação de pesos.

A função de ativação emula o limite de perceptron. Para o backpropagation você precisa de uma função continuamente diferenciável que gere um & # 8216; soft & # 8217; passo com um certo valor x. Normalmente, é utilizada uma função sigmoide, tanh ou softmax. Às vezes, também é uma função linear que apenas retorna a soma ponderada de todas as entradas. Nesse caso, a rede pode ser usada para regressão, para prever um valor numérico em vez de um resultado binário.

As redes neurais estão disponíveis na instalação R padrão (nnet, uma única rede de camada oculta) e em muitos pacotes, por exemplo RSNNS e FCNN4R.

6. Aprendizagem profunda.

Métodos de aprendizado profundo usam redes neurais com muitas camadas ocultas e milhares de neurônios, que não podem ser treinados de forma efetiva por backpropagation convencional. Vários métodos tornaram-se populares nos últimos anos para treinar tais redes enormes. Eles costumam pré-treinar as camadas do neurônio escondido para alcançar um processo de aprendizagem mais eficaz. Uma Máquina Boltzmann Restrita (RBM) é um algoritmo de classificação não supervisionado com uma estrutura de rede especial que não possui conexões entre os neurônios ocultos. Um auto-codificador esparso (SAE) usa uma estrutura de rede convencional, mas pré-treina as camadas ocultas de forma inteligente, reproduzindo os sinais de entrada nas saídas da camada com o menor número possível de conexões ativas. Esses métodos permitem redes muito complexas para lidar com tarefas de aprendizagem muito complexas. Como bater o melhor jogador humano do mundo.

As redes de aprendizagem profunda estão disponíveis nos pacotes Deepnet e Darch R. Deepnet fornece um autoencoder, Darch uma máquina Boltzmann restrito. Eu ainda não experimentei com o Darch, mas aqui é um exemplo de script R usando o autoencoder Deepnet com 3 camadas ocultas para sinais comerciais através da função neural () do Zorro & # 8217;

7. Suporte máquinas vetoriais.

Como uma rede neural, uma máquina de vetor de suporte (SVM) é outra extensão da regressão linear. Quando olhamos novamente para a fórmula de regressão,

podemos interpretar os recursos x n como coordenadas de um espaço de recursos n-dimensional. Definir a variável de destino y para um valor fixo determina um plano nesse espaço, chamado de hiperplane, pois possui mais de duas dimensões (na verdade, n-1). O hiperplane separa as amostras com y & gt; o das amostras com y & lt; 0. Os coeficientes a n podem ser calculados de forma a que as distâncias do plano para as amostras mais próximas # 8211; que são chamados de & # 8216; vetores de suporte & # 8217; do plano, daí o nome do algoritmo & # 8211; é o máximo. Desta forma, temos um classificador binário com a separação ideal de amostras vencedoras e perdidas.

O problema: normalmente, essas amostras não são linearmente separáveis ​​e # 8211; Eles estão espalhados irregularmente no espaço de recursos. Nenhum avião plano pode ser espremido entre vencedores e perdedores. Se pudesse, tínhamos métodos mais simples para calcular esse avião, f. i. análise discriminante linear. Mas, no caso comum, precisamos do truque SVM: adicionando mais dimensões ao espaço de recursos. Para isso, o algoritmo SVM produz mais recursos com uma função kernel que combina dois preditores existentes para um novo recurso. Isso é análogo ao passo acima, desde a regressão simples até a regressão polinomial, onde também são adicionados mais recursos, levando o único preditor ao n-ésimo poder. Quanto mais dimensões você adiciona, mais fácil é separar as amostras com um hiperplano plano. Este plano é então transformado de volta para o espaço n-dimensional original, ficando enrugado e amassado no caminho. Através da seleção inteligente da função kernel, o processo pode ser executado sem realmente calcular a transformação.

À semelhança das redes neurais, os SVMs podem ser utilizados não apenas para classificação, mas também para regressão. Eles também oferecem alguns parâmetros para otimizar e possivelmente superar o processo de previsão:

Função Kernel. Você normalmente usa um kernel RBF (função de base radial, um kernel simétrico), mas você também tem a escolha de outros kernels, como sigmoid, polynomial e linear. Gamma, a largura do kernel RBF Custo parâmetro C, & # 8216; penalidade & # 8217; para classificações erradas nas amostras de treino.

Um SVM usado frequentemente é a biblioteca libsvm. Ele também está disponível em R no pacote e1071. Na próxima e última parte desta série, planejo descrever uma estratégia comercial usando este SVM.

8. K-vizinho mais próximo.

Comparado com as coisas pesadas de ANN e SVM, esse é um bom algoritmo simples com uma propriedade única: não precisa de treinamento. Então as amostras são o modelo. Você poderia usar esse algoritmo para um sistema comercial que aprenda permanentemente simplesmente adicionando mais e mais amostras. O algoritmo vizinho mais próximo calcula as distâncias no espaço de recursos dos valores de recurso atuais para as amostras mais próximas do k. Uma distância no espaço n-dimensional entre dois conjuntos de recursos (x 1 .. x n) e (y 1 .. y n) é calculada exatamente como em 2 dimensões:

O algoritmo simplesmente prediz o alvo da média das k variáveis ​​alvo das amostras mais próximas, ponderadas por suas distâncias inversas. Pode ser usado para classificação, bem como para regressão. Os truques de software emprestados a partir de gráficos de computador, como uma árvore binária adaptativa (ABT), podem fazer com que o vizinho mais próximo busque muito rápido. Na minha vida passada como programador de jogos de computador, usamos esses métodos em jogos para tarefas como inteligência inimiga de auto-aprendizagem. Você pode chamar a função knn em R para a previsão do vizinho mais próximo e # 8211; ou escreva uma função simples em C para esse propósito.

Este é um algoritmo de aproximação para classificação não supervisionada. Tem alguma semelhança, não apenas com o nome, com o vizinho mais próximo. Para classificar as amostras, o algoritmo primeiro coloca k pontos aleatórios no espaço de recursos. Em seguida, atribui a qualquer um desses pontos todas as amostras com as menores distâncias a ele. O ponto é então movido para a média dessas amostras mais próximas. Isso gerará uma nova atribuição de amostras, uma vez que algumas amostras estão agora mais próximas de outro ponto. O processo é repetido até a atribuição não mudar mais movendo os pontos, isto é, cada ponto está exatamente na média das amostras mais próximas. Agora temos k classes de amostras, cada uma na vizinhança de um dos pontos k.

Este algoritmo simples pode produzir resultados surpreendentemente bons. Em R, a função kmeans faz o truque. Um exemplo do algoritmo k-means para classificar padrões de velas pode ser encontrado aqui: classificação de castiçal não supervisionada para diversão e lucro.

10. Naive Bayes.

Este algoritmo usa Bayes & # 8217; Teorema para classificar amostras de características não numéricas (isto é, eventos), como os padrões de vela acima mencionados. Suponha que um evento X (por exemplo, que o Open da barra anterior esteja abaixo do Open da barra atual) aparece em 80% de todas as amostras vencedoras. Qual é então a probabilidade de uma amostra estar ganhando quando contém evento X? Não é 0.8 como você pensa. A probabilidade pode ser calculada com Bayes & # 8217; Teorema:

P (Y | X) é a probabilidade de que o evento Y (f. i. winning) ocorra em todas as amostras contendo evento X (no nosso exemplo, Abrir (1) & lt; Abrir (0)). De acordo com a fórmula, é igual à probabilidade de X ocorrer em todas as amostras vencedoras (aqui, 0,8), multiplicado pela probabilidade de Y em todas as amostras (cerca de 0,5 quando você seguiu meu conselho acima de amostras equilibradas) e dividido por a probabilidade de X em todas as amostras.

Se somos ingênuos e assumimos que todos os eventos X são independentes um do outro, podemos calcular a probabilidade geral de que uma amostra ganhe simplesmente multiplicando as probabilidades P (X | winning) para cada evento X. Desta forma, acabamos com esta fórmula:

com um fator de escala s. Para que a fórmula funcione, os recursos devem ser selecionados de forma que sejam o mais independentes possível, o que impõe um obstáculo ao uso de Naive Bayes na negociação. Por exemplo, os dois eventos fecham (1) & lt; Fechar (0) e Abrir (1) & lt; Open (0) provavelmente não são independentes um do outro. Os preditores numéricos podem ser convertidos em eventos dividindo o número em intervalos separados.

O algoritmo Naive Bayes está disponível no omnipresente pacote e1071 R.

11. Árvores de decisão e regressão.

Essas árvores predizem um resultado ou um valor numérico com base em uma série de decisões sim / não, em uma estrutura como os ramos de uma árvore. Qualquer decisão é a presença de um evento ou não (no caso de características não numerais) ou uma comparação de um valor de recurso com um limite fixo. Uma função de árvore típica, gerada pelo construtor de árvores do Zorro & # 8217; parece assim:

Como uma tal árvore é produzida a partir de um conjunto de amostras? Existem vários métodos; Zorro usa a entropia Shannon i nformation, que já teve uma aparição neste blog no artigo Scalping. No começo, verifica um dos recursos, digamos x 1. Coloca um hiperplano com a fórmula plana x 1 = t no espaço da característica. Este hiperplato separa as amostras com x 1 & gt; t das amostras com x 1 & lt; t. O limite de divisão t é selecionado de modo que o ganho de informação & # 8211; a diferença de entropia de informação de todo o espaço, a soma das entropias de informação dos dois sub-espaços divididos e # 8211; é o máximo. Este é o caso quando as amostras nos subespaços são mais parecidas entre si que as amostras em todo o espaço.

Este processo é então repetido com o próximo recurso x 2 e dois hiperplanos dividindo os dois subespaços. Cada divisão é equivalente a uma comparação de um recurso com um limite. Por fraccionamento repetido, logo obteremos uma enorme árvore com milhares de comparações de limiar. Em seguida, o processo é executado para trás pela poda da árvore e remoção de todas as decisões que não levam a um aumento substancial de informações. Finalmente, acabamos com uma árvore relativamente pequena como no código acima.

As árvores de decisão possuem uma ampla gama de aplicações. Eles podem produzir excelentes previsões superiores às das redes neurais ou às máquinas de vetor de suporte. Mas eles não são uma solução única, já que seus planos de divisão são sempre paralelos aos eixos do espaço de recursos. Isso limita um pouco suas previsões. Eles podem ser usados ​​não só para classificação, mas também para regressão, por exemplo, retornando a porcentagem de amostras que contribuem para um determinado ramo da árvore. A árvore do Zorro é uma árvore de regressão. O algoritmo de árvore de classificação mais conhecido é C5.0, disponível no pacote C50 para R.

Para melhorar a previsão ainda mais ou superar a limitação do eixo paralelo, um conjunto de árvores pode ser usado, chamado floresta aleatória. A previsão é então gerada pela média ou votação das previsões das árvores individuais. As florestas aleatórias estão disponíveis em pacotes R randomForest, ranger e Rborist.

Conclusão.

Existem vários métodos diferentes de mineração de dados e aprendizagem de máquinas à sua disposição. A questão crítica: o que é melhor, uma estratégia de aprendizagem baseada em modelos ou a máquina? Não há dúvida de que o aprendizado automático da máquina tem muitas vantagens. Você não precisa se preocupar com a microestrutura do mercado, a economia, a psicologia do comerciante ou coisas suaves semelhantes. Você pode se concentrar na matemática pura. O aprendizado de máquina é uma maneira muito mais elegante e atraente de gerar sistemas de comércio. Ele tem todas as vantagens do seu lado, mas um. Apesar de todos os tópicos entusiasmados nos fóruns de comerciantes, ele tende a falhar misteriosamente na negociação ao vivo.

A cada segunda semana, um novo artigo sobre comércio com métodos de aprendizagem de máquinas é publicado (alguns podem ser encontrados abaixo). Pegue todas essas publicações com um grão de sal. De acordo com alguns papéis, as taxas de ganhos fantásticos na faixa de 70%, 80% ou mesmo 85% foram alcançadas. Embora a taxa de ganhos não seja o único critério relevante & # 8211; você pode perder mesmo com uma alta taxa de vitória e # 8211; 85% de precisão na previsão de trades é normalmente equivalente a um fator de lucro acima de 5. Com esse sistema, os cientistas envolvidos devem ser bilionários enquanto isso. Infelizmente, eu nunca consegui reproduzir as taxas de vitórias com o método descrito, e nem chegou perto. Então, talvez um monte de viés de seleção tenha entrado nos resultados. Ou talvez eu seja muito estúpido.

Em comparação com as estratégias baseadas em modelos, eu não vi muitos sistemas de aprendizado de máquina bem sucedidos até agora. E do que se ouve sobre os métodos algorítmicos por hedge funds bem-sucedidos, a aprendizagem por máquinas parece ainda raramente ser usada. Mas talvez isso mude no futuro com a disponibilidade de mais poder de processamento e a próxima de novos algoritmos para aprendizagem profunda.

Classificação usando redes neurais profundas: Dixon. et. al.2018 Previsão de direção de preço usando ANN & amp; SVM: Kara. et. al.2018 Comparação empírica de algoritmos de aprendizagem: Caruana. et. al.2006 Tendência do mercado de ações de mineração com GA & amp; SVM: Yu. Wang. Lai.2005.

A próxima parte desta série tratará do desenvolvimento prático de uma estratégia de aprendizado de máquinas.

30 pensamentos sobre & ldquo; Build Better Strategies! Parte 4: Machine Learning & rdquo;

Bela postagem. Existe uma grande quantidade de potencial nessa abordagem em relação ao mercado.

Btw você está usando o editor de código que vem com zorro? como é possível obter essa configuração de cor?

O script colorido é produzido pelo WordPress. Você não pode mudar as cores no editor do Zorro, mas você pode substituí-lo por outros editores que suportem cores individuais, por exemplo Notepad ++.

É então possível que o bloco de notas detecte as variáveis ​​zorro nos scripts? Quero dizer que o BarPeriod é comentado como está com o editor zorro?

Teoricamente sim, mas para isso você precisou configurar o destaque de sintaxe do Notepad ++ e digitar todas as variáveis ​​na lista. Tanto quanto eu sei, o Notepad ++ também não pode ser configurado para exibir a descrição da função em uma janela, como faz o editor Zorro. Não existe uma ferramenta perfeita e # 8230;

Conforme o último parágrafo. Eu tentei muitas técnicas de aprendizado de máquina depois de ler vários & # 8216; peer reviewed & # 8217; papéis. Mas reproduzir seus resultados permanece indescritível. Quando eu vivo teste com ML, eu não posso parecer melhorar a entrada aleatória.

ML falha ao vivo? Talvez o treinamento do ML tenha que ser feito com dados de preços que incluam também o spread histórico, roll, tick e assim por diante?

Eu acho que o motivo # 1 para falha ao vivo é o viés de mineração de dados, causado por seleção tendenciosa de entradas e parâmetros para o algo.

Obrigado ao autor pela grande série de artigos.

No entanto, deve-se notar que não precisamos restringir nossa visão ao prever apenas o próximo movimento de preços. Pode acontecer que o próximo movimento vá contra o nosso comércio em 70% dos casos, mas ainda vale a pena fazer um comércio. Isso acontece quando o preço finalmente vai para a direção certa, mas antes disso pode fazer alguns passos contra nós. Se atrasarmos o comércio por um passo de preço, não entraremos nos 30% mencionados das negociações, mas para isso aumentamos o resultado do passo de preço de 70% por um preço. Portanto, o critério é qual o valor mais alto: N * average_result ou 0.7 * N * (avergae_result + price_step).

Bela postagem. Se você quiser apenas brincar com alguma aprendizagem de máquinas, implementei uma ferramenta ML muito simples em python e adicionei uma GUI. É implementado para prever séries temporais.

Obrigado JCL Achei muito interessante o seu artigo. Gostaria de perguntar-lhe, a partir da sua experiência em negociação, onde podemos transferir dados históricos confiáveis ​​de forex? Eu considero isso muito importante devido ao fato de o mercado Forex estar descentralizado.

Desde já, obrigado!

Não há dados de Forex realmente confiáveis, uma vez que todo corretor de Forex cria seus próprios dados. Todos eles diferem ligeiramente dependentes de quais provedores de liquidez eles usam. FXCM tem relativamente bom M1 e marca dados com poucas lacunas. Você pode baixá-lo com o Zorro.

Obrigado por escrever uma série tão grande de artigos JCL & # 8230; uma leitura completamente agradável!

I have to say though that I don’t view model-based and machine learning strategies as being mutually exclusive; I have had some OOS success by using a combination of the elements you describe.

To be more exact, I begin the system generation process by developing a ‘traditional’ mathematical model, but then use a set of online machine learning algorithms to predict the next terms of the various different time series (not the price itself) that are used within the model. The actual trading rules are then derived from the interactions between these time series. So in essence I am not just blindly throwing recent market data into an ML model in an effort to predict price action direction, but instead develop a framework based upon sound investment principles in order to point the models in the right direction. I then data mine the parameters and measure the level of data-mining bias as you’ve described also.

It’s worth mentioning however that I’ve never had much success with Forex.

Anyway, best of luck with your trading and keep up the great articles!

Thanks for posting this great mini series JCL.

I recently studied a few latest papers about ML trading, deep learning especially. Yet I found that most of them valuated the results without risk-adjusted index, i. e., they usually used ROC curve, PNL to support their experiment instead of Sharpe Ratio, for example.

Also, they seldom mentioned about the trading frequency in their experiment results, making it hard to valuate the potential profitability of those methods. Por que é que? Do you have any good suggestions to deal with those issues?

ML papers normally aim for high accuracy. Equity curve variance is of no interest. This is sort of justified because the ML prediction quality determines accuracy, not variance.

Of course, if you want to really trade such a system, variance and drawdown are important factors. A system with lower accuracy and worse prediction can in fact be preferable when it’s less dependent on market condictions.

“In fact the most popular – and surprisingly profitable – data mining method works without any fancy neural networks or support vector machines.”

Would you please name those most popular & surprisingly profitable ones. So I could directly use them.

I was referring to the Indicator Soup strategies. For obvious reasons I can’t disclose details of such a strategy, and have never developed such systems myself. We’re merely coding them. But I can tell that coming up with a profitable Indicator Soup requires a lot of work and time.

Well, i am just starting a project which use simple EMAs to predict price, it just select the correct EMAs based on past performance and algorithm selection that make some rustic degree of intelligence.

Jonathan. orrego@gmail offers services as MT4 EA programmer.

Obrigado pelo bom writeup. It in reality used to be a leisure account it.

Look complicated to more delivered agreeable from you!

By the way, how could we be in contact?

There are following issues with ML and with trading systems in general which are based on historical data analysis:

1) Historical data doesn’t encode information about future price movements.

Future price movement is independent and not related to the price history. There is absolutely no reliable pattern which can be used to systematically extract profits from the market. Applying ML methods in this domain is simply pointless and doomed to failure and is not going to work if you search for a profitable system. Of course you can curve fit any past period and come up with a profitable system for it.

The only thing which determines price movement is demand and supply and these are often the result of external factors which cannot be predicted. For example: a war breaks out somewhere or other major disaster strikes or someone just needs to buy a large amount of a foreign currency for some business/investment purpose. These sort of events will cause significant shifts in the demand supply structure of the FX market . As a consequence, prices begin to move but nobody really cares about price history just about the execution of the incoming orders. An automated trading system can only be profitable if it monitors a significant portion of the market and takes the supply and demand into account for making a trading decision. But this is not the case with any of the systems being discussed here.

2) Race to the bottom.

Even if (1) wouldn’t be true and there would be valuable information encoded in historical price data, you would still face following problem: there are thousands of gold diggers out there, all of them using similar methods and even the same tools to search for profitable systems and analyze the same historical price data. As a result, many of them will discover the same or very similar “profitable” trading systems and when they begin actually trading those systems, they will become less and less profitable due to the nature of the market.

The only sure winners in this scenario will be the technology and tool vendors.

I will be still keeping an eye on your posts as I like your approach and the scientific vigor you apply. Your blog is the best of its kind – keep the good work!

One hint: there are profitable automated systems, but they are not based on historical price data but on proprietary knowledge about the market structure and operations of the major institutions which control these markets. Let’s say there are many inefficiencies in the current system but you absolutely have no chance to find the information about those by analyzing historical price data. Instead you have to know when and how the institutions will execute market moving orders and front run them.

Thanks for the extensive comment. I often hear these arguments and they sound indeed intuitive, only problem is that they are easily proven wrong. The scientific way is experiment, not intuition. Simple tests show that past and future prices are often correlated – otherwise every second experiment on this blog had a very different outcome. Many successful funds, for instance Jim Simon’s Renaissance fund, are mainly based on algorithmic prediction.

One more thing: in my comment I have been implicitly referring to the buy side (hedge funds, traders etc) not to the sell side (market makers, banks). The second one has always the edge because they sell at the ask and buy at the bid, pocketing the spread as an additional profit to any strategy they might be running. Regarding Jim Simon’s Renaissance: I am not so sure if they have not transitioned over the time to the sell side in order to stay profitable. There is absolutely no information available about the nature of their business besides the vague statement that they are using solely quantitative algorithmic trading models…

Thanks for the informative post!

Regarding the use of some of these algorithms, a common complaint which is cited is that financial data is non-stationary…Do you find this to be a problem? Couldn’t one just use returns data instead which is (I think) stationary?

Yes, this is a problem for sure. If financial data were stationary, we’d all be rich. I’m afraid we have to live with what it is. Returns are not any more stationary than other financial data.

Hello sir, I developed some set of rules for my trading which identifies supply demand zones than volume and all other criteria. Can you help me to make it into automated system ?? If i am gonna do that myself then it can take too much time. Please contact me at svadukia@gmail if you are interested.

Sure, please contact my employer at info@opgroup. de. They’ll help.

I have noticed you don’t monetize your page, don’t waste your traffic,

you can earn extra bucks every month because you’ve got high quality content.

If you want to know how to make extra $$$, search for: Mrdalekjd methods for $$$

Technical analysis has always been rejected and looked down upon by quants, academics, or anyone who has been trained by traditional finance theories. I have worked for proprietary trading desk of a first tier bank for a good part of my career, and surrounded by those ivy-league elites with background in finance, math, or financial engineering. I must admit none of those guys knew how to trade directions. They were good at market making, product structures, index arb, but almost none can making money trading directions. Por quê? Because none of these guys believed in technical analysis. Then again, if you are already making your millions why bother taking the risk of trading direction with your own money. For me luckily my years of training in technical analysis allowed me to really retire after laying off from the great recession. I look only at EMA, slow stochastics, and MACD; and I have made money every year since started in 2009. Technical analysis works, you just have to know how to use it!!

KDnuggets.

Apresentamos uma lista de 50 APIs selecionadas de áreas como aprendizado de máquina, previsão, análise de texto e # 038; classificação, reconhecimento facial, tradução de idiomas, etc. Comece a consumir APIs!

Como a inteligência artificial e amp; Aprendizagem de ferramentas baseadas em aplicações evoluem, vemos vários mash ups de APIs para experimentar. Comece com esta lista de APIs selecionadas para explorar suas capacidades e amp; recursos em aprendizado de máquina, previsão, reconhecimento de rosto, processamento de imagem, reconhecimento de fala, etc. Verifique onde essas APIs são colocadas em uso!

Machine Learning for Trading.

Offered at Georgia Tech as CS 7646.

Nanodegree Program.

Machine Learning Engineer.

Make Predictive Models.

Accelerate your career with the credential that fast-tracks you to job success.

About this Course.

This course introduces students to the real world challenges of implementing machine learning based trading strategies including the algorithmic steps from information gathering to market orders. The focus is on how to apply probabilistic machine learning approaches to trading decisions. We consider statistical approaches like linear regression, KNN and regression trees and how to apply them to actual stock trading situations.

Course Cost.

Aprox. 4 months.

Skill Level.

Included in Course.

Rich Learning Content.

Taught by Industry Pros.

Student Support Community.

Join the Path to Greatness.

This free course is your first step towards a new career with the Machine Learning Engineer Nanodegree Program.

Free Course.

Machine Learning for Trading.

Enhance your skill set and boost your hirability through innovative, independent learning.

Nanodegree Program.

Machine Learning Engineer.

Accelerate your career with the credential that fast-tracks you to job success.

Course Leads.

Tucker Balch.

Arpan Chakraborty.

O que você aprenderá.

This course is composed of three mini-courses :

Mini-course 1: Manipulating Financial Data in Python Mini-course 2: Computational Investing Mini-course 3: Machine Learning Algorithms for Trading.

Each mini-course consists of about 7-10 short lessons. Assignments and projects are interleaved.

Fall 2018 OMS students : There will be two tests - one midterm after mini-course 2, and one final exam.

Prerequisites and Requirements.

Students should have strong coding skills and some familiarity with equity markets. No finance or machine learning experience is assumed.

Note that this course serves students focusing on computer science, as well as students in other majors such as industrial systems engineering, management, or math who have different experiences. All types of students are welcome!

The ML topics might be "review" for CS students, while finance parts will be review for finance students. However, even if you have experience in these topics, you will find that we consider them in a different way than you might have seen before, in particular with an eye towards implementation for trading.

Programming will primarily be in Python. We will make heavy use of numerical computing libraries like NumPy and Pandas.

Why Take This Course.

By the end of this course, you should be able to:

Understand data structures used for algorithmic trading. Know how to construct software to access live equity data, assess it, and make trading decisions. Understand 3 popular machine learning algorithms and how to apply them to trading problems. Understand how to assess a machine learning algorithm's performance for time series data (stock price data). Know how and why data mining (machine learning) techniques fail. Construct a stock trading software system that uses current daily data.

Some limitations/constraints:

We use daily data. This is not an HFT course, but many of the concepts here are relevant. We don't interact (trade) directly with the market, but we will generate equity allocations that you could trade if you wanted to.

What do I get?

Instructor videos Learn by doing exercises Taught by industry professionals.

Cursos relacionados.

Machine Learning: Unsupervised Learning.

Knowledge-Based AI: Cognitive Systems.

Health Informatics in the Cloud.

Big Data Analytics in Healthcare.

Deploying a Hadoop Cluster.

Segmentation and Clustering.

Cursos populares.

Artificial Intelligence - Deep Learning.

Display Advertising.

VR Software Development.

Featured Programs.

Only At Udacity.

Programas.

O negócio.

"Nanodegree" is a registered trademark of Udacity. &cópia de; 2018–2018 Udacity, Inc.

Udacity is not an accredited university and we don't confer degrees.

Robot Wealth.

Posted on March 4, 2018 by Kris Longmore.

Update 2 : Responding to another suggestion, I’ve added some equity curves of a simple trading system using the knowledge gained from this analysis.

Update 3 : In response to a comment from Alon, I’ve added some Lite-C code that generates the training data used in this post and outputs it to a. csv file. You can have some serious fun with this and it enables you to greatly extend the research presented here. Apreciar!

Update 4 : 30 March 2017: The analysis was extended the analysis to include data up to the end of 2018. Of course, I updated the post to reflect the changes.

The code for performing the analysis and generating the data I used in this post, as well the data file itself and another for a different instrument that you can experiment with, are all available for free download by joining our mailing list.

One of the first books I read when I began studying the markets a few years ago was David Aronson’s Evidence Based Technical Analysis . The engineer in me was attracted to the ‘Evidence Based’ part of the title. This was soon after I had digested a trading book that claimed a basis in chaos theory, the link to which actually turned out to be non-existent. Apparently using complex-sounding terms in the title of a trading book lends some measure of credibility. Anyway, Evidence Based Technical Analysis is largely a justification of a scientific approach to trading, including a method for rigorous assessment of the presence of data mining bias in backtest results. There is also a compelling discussion based in cognitive psychology of the reasons that some traders turn away from objective methods and embrace subjective beliefs. I find this area fascinating.

Readers of this blog will know that I am very interested in using machine learning to profit from the markets. Imagine my delight when I discovered that David Aronson had co-authored a new book with Timothy Masters titled Statistically Sound Machine Learning for Algorithmic Trading of Financial Instruments – which I will herein refer to as SSML. I quickly devoured the book and have used it as a handy reference ever since. While it is intended as a companion to Aronson’s (free) software platform for strategy development, it contains numerous practical tips for any machine learning practitioner and I’ve implemented most of his ideas in R.

I used SSML to guide my early forays into machine learning for trading, and this series describes some of those early experiments. While a detailed review of everything I learned from SSML and all the research it inspired is a bit voluminous to relate in detail, what follows is an account of what I found to be some of the more significant and practical learnings that I encountered along the way.

This post will focus on feature engineering and also introduce the data mining approach. The next post will focus on algorithm selection and ensemble methods for combining the predictions of numerous learners.

The data mining approach.

Data mining is just one approach to extracting profits from the markets and is different to a model-based approach. Rather than constructing a mathematical representation of price, returns or volatility from first principles, data mining involves searching for patterns first and then fitting a model to those patterns after the fact. Both model-based and data mining approaches have pros and cons, and I contend that using both approaches can lead to a valuable source of portfolio diversity.

The Financial Hacker summed up the advantages and disadvantages of the data mining approach nicely:

The advantage of data mining is that you do not need to care about market hypotheses. The disadvantage: those methods usually find a vast amount of random patterns and thus generate a vast amount of worthless strategies. Since mere data mining is a blind approach, distinguishing real patterns – caused by real market inefficiencies – from random patterns is a challenging task. Even sophisticated reality checks can normally not eliminate all data mining bias. Not many successful trading systems generated by data mining methods are known today.

David Aronson himself cautions against putting blind faith in data mining methods:

Though data mining is a promising approach for finding predictive patterns in data produced by largely random complex processes such as financial markets, its findings are upwardly biased. This is the data mining bias. Thus, the profitability of methods discovered by data mining must be evaluated with specialized statistical tests designed to cope with the data mining bias.

I would add that the implicit assumption behind the data mining approach is that the patterns identified will continue to repeat in the future. Of course, the validity of this assumption is unlikely to ever be certain.

Data mining is a term that can mean different things to different people depending on the context. When I refer to a data mining approach to trading systems development, I am referring to the use of statistical learning algorithms to uncover relationships between feature variables and a target variable (in the regression context, these would be referred to as the independent and dependent variables, respectively). The feature variables are observations that are assumed to have some relationship to the target variable and could include, for example, historical returns, historical volatility, various transformations or derivatives of a price series, economic indicators, and sentiment barometers. The target variable is the object to be predicted from the feature variables and could be the future return (next day return, next month return etc), the sign of the next day’s return, or the actual price level (although the latter is not really recommended, for reasons that will be explained below).

Although I differentiate between the data mining approach and the model-based approach, the data mining approach can also be considered an exercise in predictive modelling. Interestingly, the model-based approaches that I have written about previously (for example ARIMA, GARCH, Random Walk etc) assume linear relationships between variables. Modelling non-linear relationships using these approaches is (apparently) complex and time consuming. On the other hand, some statistical learning algorithms can be considered ‘universal approximators’ in that they have the ability to model any linear or non-linear relationship. It was not my intention to get into a philosophical discussion about the differences between a model-based approach and a data mining approach, but clearly there is some overlap between the two. In the near future I am going to write about my efforts to create a hybrid approach that attempts a synergistic combination of classical linear time series modelling and non-linear statistical learning – trust me, it is actually much more interesting than it sounds. Assista esse espaço.

Variables and feature engineering.

The prediction target.

The first and most obvious decision to be made is the choice of target variable. In other words, what are we trying to predict? For one-day ahead forecasting systems, profit is the usual target. I used the next day’s return normalized to the recent average true range, the implication being that in live trading, position sizes would be inversely proportionate to the recent volatility. In addition, by normalizing the target variable in this way, we may be able to train the model on multiple markets, as the target will be on the same scale for each.

Choosing predictive variables.

In SSML, Aronson states that the golden rule of feature selection is that the predictive power should come primarily from the features and not from the model itself. My research corroborated this statement, with many (but not all) algorithm types returning correlated predictions for the same feature set. I found that the choice of features had a far greater impact on performance than choice of model. The implication is that spending considerable effort on feature selection and feature engineering is well and truly justified. I believe it is critical to achieving decent model performance.

Many variables will have little or no relationship with the target variable and including these will lead to overfitting or other forms of poor performance. Aronson recommends using chi-square tests and Cramer’s V to quantify the relationship between variables and the target. I actually didn’t use this approach, so I can’t comment on it. I used a number of other approaches, including ranking a list of candidate features according to their Maximal Information Coefficient (MIC) and selecting the highest ranked features, Recursive Feature Elimination (RFE) via the caret package in R, an exhaustive search of all linear models, and Principal Components Analysis (PCA). Each of these are discussed below.

Some candidate features.

Following is the list of features that I investigated as part of this research. Most were derived from SSML. This list is by no means exhaustive and only consists of derivatives and transformations of the price series. I haven’t yet tested exogenous variables, such as economic indicators, the price histories of other instruments and the like, but I think these are deserving of attention too. The following list is by no means exhaustive, but provides a decent starting point:

1-day log return Trend deviation: the logarithm of the closing price divided by the lowpass filtered price Momentum: the price today relative to the price x days ago, normalized by the standard deviation of daily price changes. ATR: the average true range of the price series Velocity: a one-step ahead linear regression forecast on closing prices Linear forecast deviation: the difference between the most recent closing price and the closing price predicted by a linear regression line Price variance ratio: the ratio of the variance of the log of closing prices over a short time period to that over a long time period. Delta price variance ratio: the difference between the current value of the price variance ratio and its value x periods ago. The Market Meanness Index: A measure of the likelihood of the market being in a state of mean reversion, created by the Financial Hacker. MMI deviation: The difference between the current value of the Market Meanness Index and its value x periods ago. The Hurst exponenet ATR ratio: the ratio of an ATR of a short (recent) price history to an ATR of a longer period. Delta ATR ratio: the difference between the current value of the ATR ratio and the value x bars ago. Bollinger width: the log ratio of the standard deviation of closing prices to the mean of closing prices, that is a moving standard deviation of closing prices relative to the moving average of closing prices. Delta bollinger width: the difference between the current value of the bollinger width and its value x bars ago. Absolute price change oscillator: the difference between a short and long lookback mean log price divided by a 100-period ATR of the log price.

Thus far I have only considered the most recent value of each variable. I suspect that the recent history of each variable would provide another useful dimension of data to mine. I left this out of the feature selection stage since it makes more sense to firstly identify features whose current values contain predictive information about the target variable before considering their recent histories. Incorporating this from the beginning of the feature selection stage would increase the complexity of the process by several orders of magnitude and would be unlikely to provide much additional value. I base that statement on a number of my own assumptions, not to mention the practicalities of the data mining approach, rather than any hard evidence.

Transforming the candidate features.

In my experiments, the variables listed above were used with various cutoff periods (that is, the number of periods used in their calculation). Typically, I used values between 3 and 20 since Aronson states in SSML that lookback periods greater than about 20 will generally not contain information useful to the one period ahead forecast. Some variables (like the Market Meanness Index) benefit from a longer lookback. For these, I experimented with 50, 100, and 150 bars.

Additionally, it is important to enforce a degree of stationarity on the variables. Davind Aronson again:

Using stationary variables can have an enormous positive impact on a machine learning model. There are numerous adjustments that can be made in order to enforce stationarity such as centering, scaling, and normalization. So long as the historical lookback period of the adjustment is long relative to the frequency of trade signals, important information is almost never lost and the improvements to model performance are vast.

Scaling: divide the indicator by the interquartile range (note, not by the standard deviation, since the interquartile range is not as sensitive to extremely large or small values). Centering: subtract the historical median from the current value. Normalization: both of the above. Roughly equivalent to traditional z-score standardization, but uses the median and interquartile range rather than the mean and standard deviation in order to reduce the impact of outliers. Regular normalization: standardizes the data to the range -1 to +1 over the lookback period (x-min)/(max-min) and re-centered to the desired range.

In my experiments, I generally adopted regular normalization using the most recent 50 values of the features.

Data Pre-Processing.

If you’re following along with the code and data provided (see note in bold above), I used the data for the GBP/USD exchange rate (sampled daily at midnight UTC, for the period 2009-2018), but I also provided data for EUR/USD (same sampling regime) for further experimentation.

Removing highly correlated variables.

It makes sense to remove variables that are highly correlated with other variables since they are unlikely to provide additional information that isn’t already contained elsewhere in the feature space. Keeping these variables will also add unnecessary computation time, increase the risk of overfitting and bias the final model towards the correlated variables.

Running caret ‘s function for examining pairwise correlations between variables – caret :: findCorrelation ( ) – with a cutoff of 0.3, these are the remaining variables and their pairwise correlations:

Feature selection via Maximal Information.

The maximal information coefficient (MIC) is a non-parametric measure of two-variable dependence designed specifically for rapid exploration of many-dimensional data sets. While MIC is limited to univariate relationships (that is, it does not consider variable interactions), it does pick up non-linear relationships between dependent and independent variables. Read more about MIC here. I used the minerva package in R to rank my variables according to their MIC with the target variable (next day’s return normalized to the 100-period ATR). Here’s the code output:

These results show that none of the features have a particularly high MIC with respect to the target variable, which is what I would expect from noisy data such as daily exchange rates sampled at an arbitrary time. However, certain variables have a higher MIC than others. In particular, the long-term ATR and the 20-period Hurst exponent and the 3-period Bollinger width outperform the rest of the variables.

Recursive feature elimination.

I also used recursive feature elimination (RFE) via the caret package in R to isolate the most predictive features from my list of candidates. RFE is an iterative process that involves constructing a model from the entire set of features, retaining the best performing features, and then repeating the process until all the features are eliminated. The model with the best performance is identified and the feature set from that model declared the most useful.

I performed cross-validated RFE using a random forest model. Aqui estão os resultados:

In this case, the RFE process has emphasized variables that describe volatility and trend, but has decided that the best performance is obtained by incorporating 14 of 15 variables into the model. Here’s a plot of the cross validated performance of the best feature set for various numbers of features (noting that k - fold cross validation may not be the ideal cross-validation method for financial time series):

I am tempted to take the results of the RFE with a grain of salt. My reasons are:

The RFE algorithm does not fully account for interactions between variables. For example, assume that two variables individually have no effect on model performance, but due to some relationship between them they improve performance when both are included in the feature set. RFE is likely to miss this predictive relationship. The performance of RFE is directly related to the ability of the specific algorithm (in this case random forest) to uncover relationships between the variables and the target. At this stage of the process, we have absolutely no evidence that the random forest model is applicable in this sense to our particular data set. Finally, the implementation of RFE that I used was the ‘out of the box’ caret version. This implementation uses root mean squared error (RMSE) as the objective function, however I don’t believe that RMSE is the best objective function for this data due to the significant influence of extreme values on model performance. It is possible to have a low RMSE but poor overall performance if the model is accurate across the middle regions of the target space (corresponding to small wins and losses), but inaccurate in the tails (corresponding to big wins and losses)

In order to address (3) above, I implemented a custom summary function so that the RFE was performed such that the cross-validated absolute return was maximized. I also applied the additional criterion that only predictions with an absolute value of greater than 5 would be considered under the assumption that in live trading we wouldn’t enter positions unless the prediction exceeded this value. Os resultados são os seguintes:

The results are a little different to those obtained using RMSE as the objective function. The focus is still on the volatility and trend indicators, but in this case the best cross validated performance occurred when selecting only 2 out of the 15 candidate variables. Here’s a plot of the cross validated performance of the best feature set for various numbers of features:

The model clearly performs better in terms of absolute return for a smaller number of predictors. This is consistent with Aronson’s assertion that with this approach we should stick with at most 3-4 variables otherwise overfitting is almost unavoidable.

The performance profile of the model tuned on absolute return is very different to that of the model tuned on RMSE, which displays a consistent improvement as the number of predictors is increased. Using RMSE as the objective function (which seems to be the default in many applications I’ve come across) would result in a very sub-optimal final model in this case. This highlights the importance of ensuring that the objective function is a good proxy for the performance being sought in practice.

In the RFE example above, I used 5-fold cross validation, but I haven’t held out a test set of data or estimated performance with an inner cross validation loop. Note also that k - fold cross validation may not be ideal for financial time series thanks to the autocorrelations present.

Models with in-built feature selection.

A number of machine learning algorithms have feature selection in-built. Max Kuhn’s website for the caret package contains a list of such models that are accessible through the caret package. I’ll apply several and compare the features selected to those selected with other methods. For this experiment, I used a diverse range of algorithms that include various ensemble methods and both linear and non-linear interactions:

Bagged multi-adaptive regressive splines (MARS) Boosted generalized additive model (bGAM) Lasso Spike and slab regression (SSR) Regression tree Stochastic gradient boosting (SGB)

For each model, I did only very basic (if any) hyperparameter tuning within caret using time series cross validation with a train window length of 200 days and a test window length of 20 days. Maximization of absolute return was used as the objective function. Following cross-validation, caret actually trains a model on the full data set with the best cross-validated hyperparameters – but this is not what we want if we are to mimic actual trading behaviour (we are more interested in the aggregated performance across each test window, which caret very neatly allows us to access – details on this below when we investigate a trading system).

The table below shows the top 5 variables for each algorithm:

We can see that 10-day momentum is included in the top 5 predictors for every algorithm I investigated except one, and was the top feature every time it was selected. The change in the ratio of the ATR lookbacks featured 7 times in total. A responsive absolute price change oscillator was selected 4 times, and in one form or another (10- and 20-day variables once each and 100-day variable thrice). The 5-day change in the price variance ratio was a notable mention, being included in the top variables 3 times. The table below summarizes the frequency with which each variable was selected:

9 of the 15 variables that passed the correlation filter were selected in the top 5 by at least one algorithm.

Model selection using glmulti.

The glmulti package fits all possible unique generalized linear models from the variables and returns the ‘best’ models as determined by an information criterion (Aikake in this case). The package is essentially a wrapper for the glm (generalized linear model) function that allows selection of the ‘best’ model or models, providing insight into the most predictive variables. By default, glmulti builds models from the main interactions, but there is an option to also include pairwise interactions between variables. This increases the computation time considerably, and I found that the resulting ‘best’ models were orders of magnitude more complex than those obtained using main interactions only, and results were on par.

1 + trend + atrRatSlow"

We retain the models whose AICs are less than two units from the ‘best’ modelo. Two units is a rule of thumb for models that, for all intents and purposes, are likely to be on par in terms of their performance:

1 + trend + atrRatSlow 22669.87 0.03235620.

1 + apc5 + trend + atrRatSlow 22670.72 0.02111157.

1 + MMIFaster + trend + atrRatSlow 22670.73 0.02104022.

1 + bWdith3 + trend + atrRatSlow 22670.87 0.01961736.

1 + trend + atrRatSlow + bWidthSlow 22670.96 0.01878187.

1 + deltaMMIFastest10 + trend + atrRatSlow 22671.04 0.01799825.

1 + trend + atrRatSlow + ATRSlow 22671.20 0.01667710.

1 + atrRatSlow 22671.23 0.01644049.

1 + apc5 + MMIFaster + trend + atrRatSlow 22671.40 0.01504698.

1 + deltaPVR5 + trend + atrRatSlow 22671.44 0.01474030.

1 + MMIFaster + trend + atrRatSlow + bWidthSlow 22671.49 0.01439135.

1 + HurstFast + trend + atrRatSlow 22671.55 0.01396476.

1 + HurstMod + trend + atrRatSlow 22671.61 0.01357259.

1 + mom10 + trend + atrRatSlow 22671.64 0.01337186.

1 + apc5 + bWdith3 + trend + atrRatSlow 22671.65 0.01329943.

1 + MMIFaster + deltaMMIFastest10 + trend + atrRatSlow 22671.66 0.01325762.

1 + deltaATRrat3 + trend + atrRatSlow 22671.68 0.01309016.

1 + deltaATRrat10 + trend + atrRatSlow 22671.69 0.01303342.

1 + HurstFaster + trend + atrRatSlow 22671.70 0.01294480.

1 + bWdith3 + MMIFaster + trend + atrRatSlow 22671.72 0.01283320.

1 + apc5 + trend + atrRatSlow + bWidthSlow 22671.74 0.01268352.

1 + apc5 + deltaMMIFastest10 + trend + atrRatSlow 22671.83 0.01216110.

1 + bWdith3 + trend + atrRatSlow + bWidthSlow 22671.85 0.01205142.

Notice any patterns here? many of the top models selected the ratio of the 20- day to 100-day ATRs, as well as the difference between a short-term and long-term trend indicator. Perhaps surprisingly sparse are the momentum variables. This is confirmed with this plot of the model averaged variable importance (averaged over the best 1,000 models):

Note that these models only considered the main, linear interactions between each variable and the target. Of course, there is no guarantee that any relationship is linear, if it exists at all. Further, there is the implicit assumption of stationary relationships amongst the variables which is unlikely to hold. Still, this method provides some useful insight.

One of the great things about glmulti is that it facilitates model-averaged predictions – more on this when I delve into ensembles in part 2 of this series.

Generalized linear model with stepwise feature selection.

Finally, I used a generalized linear model with stepwise feature selection:

The final model selected 2 of the 15 variables: the ratio of the 20- to 100-day ATR, difference between a short-term and long-term trend indicator.

Boruta: all relevant feature selection.

Boruta finds relevant features by comparing the importance of the original features with the importance of random variables. Random variables are obtained by permuting the order of values of the original features. Boruta finds a minimum, mean and maximum value of the importance of these permuted variables, and then compares these to the original features. Any original feature that is found to be more relevant than the maximum random permutation is retained.

Boruta does not measure the absolute importance of individual features, rather it compares each feature to random permutations of the original variables and determines the relative importance. This theory very much resonates with me and I intuit that it will find application in weeding out uninformative features from noisy financial data. The idea of adding randomness to the sample and then comparing performance is analogous to the approach I use to benchmark my systems against a random trader with a similar trade distribution.

The box plots in the figure below show the results obtained when I ran the Boruta algorithm for the 15 filtered variables for 1,000 iterations. The blue box plots show the permuted variables of minimum, mean and maximum importance, the green box plots indicate the original features that ranked higher than the maximum importance of the random permuted variables, and the variables represented by the red box plots are discarded.

These results are largely consistent with the results obtained through other methods, perhaps with the exception of the inclusion of the MMI and Hurst variables. Surprisingly, the long-term ATR was the clear winner.

Side note: The developers state that “Boruta” means “Slavic spirit of the forest.” As something of a slavophile myself, I did some googling and discovered that this description is quite a euphemism. Check out some of the items that pop up in a Google image search!

Discussion of feature selection methods.

It is important to note that any feature selection process naturally invites a degree of selection bias. For example, from a large set of uninformative variables, a small number may randomly correlate with the target variable. The selection algorithm would then rank these variables highly. The error would only be (potentially) uncovered through cross validation of the selection algorithm or by using an unseen test or validation set. Feature selection is difficult and can often make predictive performance worse, since it is easy to over-fit the feature selection criterion. It is all to easy to end up with a subset of attributes that works really well on one particular sample of data, but not necessarily on any other. There is a fantastic discussion of this at the Statistics Stack Exchange community that I have linked here because it is just so useful.

It is critical to take steps to minimize selection bias at every opportunity. The results of any feature selection process should be cross validated or tested on an unseen hold out set. If the hold out set selects a vastly different set of predictors, something has obviously gone wrong – or the features are worthless. The approach I took in this post was to cross validate the results of each test that I performed, with the exception of the Maximal Information Criterion and glmulti approaches. I’ve also selected features based on data for one market only. If the selected features are not robust, this will show up with poor performance when I attempt to build predictive models for other markets using these features.

I think that it is useful to apply a wide range of methods for feature selection, and then look for patterns and consistencies across these methods. This approach seems to intuitively be far more likely to yield useful information than drawing absolute conclusions from a single feature selection process. Applying this logic to the approach described above, we can conclude that the 10-day momentum, the ratio of the 10- to 20-day ATR, the trend deviation indicator, and the absolute price change oscillator are probably the most likely to yield useful information since they continually show up in most of the feature selection methods that I investigated. Other variables that may be worth considering include the long-term ATR and the change in a responsive MMI.

In part 2 of this article, I’ll describe how I built and combined various models based on these variables.

Principal Components Analysis.

An alternative to feature selection is Principal Components Analysis (PCA), which attempts to reduce the dimensionality of the data while retaining the majority of the information. PCA is a linear technique: it transforms the data by linearly projecting it onto a lower dimension space while preserving as much of its variation as possible. Another way of saying this is that PCA attempts to transform the data so as to express it as a sum of uncorrelated components.

Again, note that PCA is limited to a linear transformation of the data, however there is no guarantee that non-linear transformations won’t be better suited. Another significant assumption when using PCA is that the principal components of future data will look those of the training data. It’s also possible that the smallest component, describing the least variance, is also the only one carrying information about the target variable, and would likely be lost when the major variance contributors are selected.

To investigate the effects of PCA on model performance, I cross validated 2 random forest models, the first using the principal components of the 15 variables, and the other using all 15 variables in their raw form. I chose the random forest model since it includes feature selection and thus may reveal some insights about how PCA stacks up in relation to other feature selection methods. For both models, I performed time series cross validation on a training window of 200 days and a testing window of 20 days.

In order to infer the difference in model performance, I collected the results from each resampling iteration of both final models and compared their distributions via a pair of box and whisker plots:

The model built on the raw data outperforms the model built on the data’s principal components in this case. The mean profit is higher and the distribution is shifted in the positive direction. Sadly however, both distributions look only sightly better than random and have wide distributions.

A simple trading system.

I will go into more detail about building a practical trading system using machine learning in the next post, but the following demonstrates a simple trading system based on some of the information gained from the analysis presented above. The system is based on three of the indicators that the feature selection analysis identified as being predictive of the target variable. The features used were long-term ATR, the change in the responsive MMI, and the trend deviance indicator. I trained a generalized boosted regression model using the gbm package in R using these indicators as the independent variables predicting the next day return normalized to the recent ATR. The model was trained on a sliding window of 200 days and tested on the adjacent 20 days over the length of the entire data set. I also subtracted transaction costs of 2 pips per round turn lot traded.

The code that was used to perform these simulations is part of an exceptionally useful research environment that enables rapid experimentation with different variables, hyperparameter tuning, and various training-testing window setups. It even enables the user to swap out the GBM algorithm for dozens of others, like neural networks, support vector machines and decision trees, with a single line of code. If you would like to make use of this code, download it for free here.

Robot Wealth members have access to the full research environment as well as other exclusive content, development tools, the members’ forum and our fundamental and advanced algo trading courses. Join the Robot Wealth community today and take your trading to the next level – sign up here.

The returns series of most financial instruments consists of a relatively large number of small positive and small negative values and a smaller number of large positive and large negative values. I hypothesize that the values whose magnitude is smaller are more random in nature than the values whose magnitude is large. On any given day, all things being equal, a small negative return could turn out to be a small positive return by the time the close rolls around, or vice versa, as a result of any number of random occurrences related to the fundamentals of the exchange rate. These same random occurrences are less likely to push a large positive return into negative territory and vice versa, purely on account of the size of the price swings involved. Following this logic, I hypothesize that my model is likely to be more accurate in its extreme predictions than in its ‘normal’ range. We can test this hypothesis on the simple trading strategy described above by entering positions only when the model predicts a return that is large in magnitude.

Here are the results, along with the buy and hold return from the testing data set:

While the strategy may look enticing, it is actually not overly robust to changing the random initialization or the hyperparameters of the GBM algorithm – more work is needed to turn this into a viable strategy. We can see that increasing the prediction threshold for entering a trade resulted in smoother equity curves, but a much reduced final equity. The model significantly outperformed the buy and hold strategy. The prediction threshold can be adjusted depending on the trader’s appetite for high returns (threshold = 0) as opposed to minimal drawdown (threshold = 100).

Conclusões.

The MIC analysis and the Boruta algorithm agreed that the long-term ATR was the most important feature. They were also the only approaches that included the Hurst exponent in their most important features. The RFE analysis indicated that it may be prudent to focus on variables that measure long-term volatility or recent changes in volatility relative to the longer term. An exhaustive search of all possible generalized linear models that considered main interactions using glmulti implied that the 20- to 100-day ATR ratio and the trend deviance variables are most predictive. Stepwise feature selection using a generalized linear model returned similar results. Boruta identified 8 useful variables, with the long-term ATR the clear winner. Transforming the variables using PCA reduced the performance of a random forest model relative to using the raw variables.

The same features seem to be selected over and over again using the different methods. Is this just a fluke, or has this long and exhaustive data mining exercise revealed something of practical use to a trader? In part 2 of this series, I will investigate the performance of various machine learning algorithms based on this feature selection exercise. I’ll compare different algorithms and then investigate combining their predictions using ensemble methods with the objective of creating a useful trading system.

Referências.

Aronson, D. 2006, Evidence-Based Technical Analysis: Applying the Scientific Method and Statistical Inference to Trading Signals .

Aronson, D. and Masters, T. 2017-, Statistically Sound Machine Learning for Algorithmic Trading of Financial Instruments: Developing Predictive-Model-Based Trading Systems Using TSSB.

Ótimo post! I’ll be sure to follow the next parts of the series. I’m just starting in the world of algo trading and based on what you wrote SSML looks like a good starting point. Would you recommend any other book?

Kris Longmore.

Obrigado! Gostou de gostar.

My reading list for starting out with algo trading includes both of Ernie Chan’s books and Aronson’s first book, ‘Evidence Based Technical Analysis’. Jaekle and Tomasini is a handy reference too. If you don’t already have the background, an introductory statistics/probability text (sorry, can’t recommend one off the top of my head). If you want to delve into time series analysis, Ruey Tsay’s book on the topic is heavy going, but worth persisting with. In terms of machine learning, Lantz’s ‘Machine Learning with R’ will get you started. Depending on your stats background, Tibrishani et. al. released two works – ‘The Elements of Statistical Learning’ (more advanced) and ‘Introduction to Statistical Learning’ & # 8211; these are the logical next steps for machine learning. Finally, I find the caret package in R to be immensely useful, and the author of the package also wrote a book called ‘Applied Predictive Modelling’ & # 8211; definitely worth a read if you are interested in the machine learning side of things.

Jacques Joubert.

Brilliant post Kris, Thank you.

This has really helped me to get started in the area of using ML in trading. Was even part of the recommended reading I was given.

Will email you about adding your recommended books here to the Quantocracy book list.

I am really looking forward to the follow up post.

Kris Longmore.

Thanks for that feedback, Jaques. I am really glad you found it useful.

I love Aronson’s books also. Did you implement a version of the MCPT ?

Kris Longmore.

Thanks Nick! I haven’t yet implemented MCPT in R, but I think the ‘coin’ package could be used. Do you know of any others? MCPT would be a good inclusion as an update to this post, or perhaps in the next.

I’m not familiar with ‘coin’ but thanks for mentioning that … I’ve implemented my own version of the MCPT.

BTW, you mention that you use Zorro. Have you used their R bridge for strategy development or live trading ?

Kris Longmore.

Yes I’ve used Zorro’s R bridge extensively for strategy development and am using it for live trading right now. Its an integral part of my workflow and I am much more efficient during the development stage compared with using either Zorro or R separately. Its a very useful piece of kit to have in your toolbox.

I ‘ll do a post in the near future on exactly how this works and why it is efficient.

Very interesting work. You might also be interested in the “Boruta” all relevant feature selection method, which can be found at m2.icm. edu. pl/boruta/

I look forward to your next few posts.

Kris Longmore.

Thanks for pointing me to Boruta, Dekalog. I’ll update the post to include this method. I see that it is implemented as an R package – how convenient!

Kris Longmore.

I’ve updated the post with a paragraph on Boruta. Thanks for pointing me to this package – very easy to use and wonderfully intuitive logic behind it.

Hi and thanks for sharing your work. One suggestion and request I would make would be to plot equity curves of some of the more promising candidates against say a simple Buy and Hold model. For example, if your in sample training set found ATR and BB (with some parameter set) useful and you created a model built around those features, you could plot both the in and out of sample resulting equity curves of the results (you also need to determine how you want to translate those predictions to actions. e. g. Buy/Sell/Flat). I’ve found the equity curve can reveal a lot about the results that may not be evident from looking at standalone metrics. One test may show promising results around a certain set of variables, but depending on how the individual results are finally aggregated, a different set of features may have performed much better.

Kris Longmore.

Hi IT, thanks for that comment. I take your point – equity curves certainly reveal a lot more useful information than the result of a feature selection test. And after all, this is a blog about trading. I had originally intended to leave all of the strategy development stuff until the next post where I will write about how I transform a model output into actionable intelligence, combining predictions from multiple learners and all that. But you’re right – an illustrative trading strategy and equity curve would fit nicely here. I’ll update the post accordingly.

Thanks again for your feedback – always much appreciated.

Deshawn Janas.

\nThat will be the common answer you will come across for every machine learning and knowledge discovery problem you face. There is no best learning system, classification system, or statistical method. You need to know your data-set well enough to realize where one approach might be better than others.

Kris Longmore.

Couldn’t agree more, Deshawn. There is an important place for subject matter knowledge in nearly any machine learning task.

Hello and thank you for the great post. I discovered TSSB thanks to it and now I’m playing with the application. I see you also use ATR normalized return. I tried to compute it in R and compare the values to TSSB outputs but they differ. For example on D1 GBP/USD data from 2009-11-11 to 2018-03-01 mean value of my computed ATR 250 normalized return is -0.01237755 while TSSB values give mean of -0.01529271. Do you experience the same (which would probably mean that the implementations differ)? Or did I make a mistake?

Kris Longmore.

Hello and thanks for commenting. Are you sampling your data at the same time in both applications? Since foreign exchange doesn’t have an official closing time, make sure that the one you use is consistent across both applications. Also, I believe TSSB calculates this variable using open prices, which in theory shouldn’t differ from the previous day’s closing price. Unless of course there is a price gap between one day’s close and the next’s open, which can happen due to weekends and holidays. Can those ideas account for the differences you are seeing?

I have not actually run anything in TSSB myself. I like using R too much!

Hey and sorry for the late reply, I was without my laptop for a while.

I completely understand, it’s easy to fall in love with R!:) I also wanted to rewrite the stuff into R but after reading the whole book I decided to give TSSB a go because I saw how much work had already been put into it.

Or maybe I will combine the approaches somehow. I noticed there is an automation framework for TSSB written in Python which could be modified to handle both: R and TSSB (unfortunately it’s Windows only solution).

About the problem – it turned out it was only because at the end of TSSB returns vector there was bunch of NA values while my computed one had all of them. I should have checked the end of the data frame more carefully, stupid mistake.

Where is “here is an automation framework for TSSB “? Perdi alguma coisa?

Hi Robot Master, great post! I was wondering if you thought about potential implication of using ATR normalized returns in your target, especially for feature selection. What I mean with that is; by normalizing with ATR you are introducing a strong recent volatility component in your target variable. And as any quant would appreciate volatility clusters heavily and therefore recent volatility measures popping-up in your best features across the board. (Maybe not only the time you redefined your target metric as profit where you only observed 2 variables.) And of course you can estimate volatility pretty decently utilizing some recent vola information but when it come to making a directional bet you won’t have much of an edge. O que você acha? I just had a quick read of the post, so apologies if I am missing something. Mantenha o bom trabalho.

Kris Longmore.

Obrigado pelo comentário. You raise a very good point – by normalizing the target variable to the current ATR, a strong volatility component is introduced. This was actually a recommendation that Aronson makes in the book on which the post is based, but may not be the best solution, particularly in relation to feature selection. I’ve done some recent work using the raw first difference of the time series and found that the feature selection algorithms tend towards slightly different features under this scenario.

This suggests to me that there may be merit in using a differenced time series as the target variable during the feature selection phase, and then during the model building phase normalizing the target variable to the ATR so as to ensure volatility is included at some level in the trading model.

Thanks for commenting, great suggestion!

Mike Scirocco.

Thank you for a very thorough writeup of your very interesting work. I’m looking forward to the next installment. The recommended reading list is appreciated as well. (I hope you will consider making a separate blog entry of Recommended Reading so the books are easy to find in one entry.)

Kris Longmore.

Obrigado por comentar. Great idea regarding the recommended reading list! I will make that my next post.

Really Thanks for posting these machine learning contents!

I’m trying to follow up and apply your posting using R.

but now, I get in trouble with making data table using variables in “Some candidate features”. I can see Zorro code in second post to make selected variables.

Since I’ve never learned the Zorro, I don’t understand how you make the variables in.

“Some candidate features” .

Could you explain how did you make variables in “Some candidate features” using R?

I know this is time-consuming but I really appreciate if you explain about making variables…

Kris Longmore.

I created those variables using Zorro and then exported them for use in R. You can reproduce them using the descriptions in the section that you are referring to, whether that be in R, C, or some other language. If you get really stuck, shoot me an email using the contact form and I’ll send you an R implementation.

I’m wondering what kind of data, lev, model(input) should be put in to operate the absretSummary function?

absretSummary <- function (data, lev = NULL, model = NULL)

Great post however…

1) cross validation.

its not allowed to use cross validation for time series as simple it introduces future leak and bias. See Rob Hydman page robjhyndman/hyndsight/crossvalidation/

2) From my experience every subset of financial data will give you different features. The method which I use to find the best features is based on bootstraping and its called Neyman – Pearson method. Different features will be selected for BUY and different for SELL side so perhaps its good to split the selection.

3) I read the paper about Boruta method. I’m using filter methods based on correlation (FCBF) and mutal information (MRMR, JMI etc). Than I bootstrap those selection by Neyman-Person method. Wrapper methods (like Boruta) take too much time to compute so you can’t bootstrap them specially if you use big and many datasets.

Kris Longmore.

Thanks for reading my blog. Regarding your points:

1) I wouldn’t say its not allowed. But I agree regular cross-validation (k-fold, bootstrap, leave-one-out etc) is not suitable for financial data. I now use and recommend the time series cross-validation approach that is mentioned at the bottom of the link you provided. See my post on selecting optimal data windows for more information.

2) Thanks for the heads up re the Neyman-Pearson method. I haven’t used this, but I will look into it. I agree that different models for long and short sides may be a good idea.

3) I find Boruta to be quite useful for my applications. I also like the intuition behind the algorithm. Its only one of many possible feature selection algoirthms though and I’m sure your methods are great too.

Thanks again for reading.

ad 3) None of the methods is great or the best, just combining with Nyman-Person give you better chance that you will not end up not relevant at all feature after selection. But its just better chance…All this analysis of financial data is very slippery, after more than 5 years of applying ML algos to it my only conclusion is than only very extensive backtest on multiple data sets from different instruments can give the answer if something works better or not. You can just try to repeat your steps on series from different instrument or the same instrument but from different period and I guess you will get different results. (e. g. different features are relevant or different technique is better)

hello, i am tom. can you tell me how to use the function absretSummary in rfeControl and rfe? Obrigado. absretSummary <- function (data, lev = NULL, model = NULL)

positions 5, sign(data[, “pred”]), 0)

Kris Longmore.

This is just a custom Summary function for use with the caret train() function. The Summary function is simply the objective function for the training process; this one trains the learning algorithm towards maximizing the profit at the end of the training period. You can easily write your own for maximizing the Sharpe ratio or any other performance metric of interest.

Max Kuhn (the author of the caret package) explains the use of the summary function here. In particular, scroll down to read about the trainControl() function and alternative performance metrics.

Espero que ajude.

Ótimo artigo! really good research in all aspects.

is it possible to post the code showing how you make the training file (gu_data)?

Kris Longmore.

Obrigado por ler! I’ve added the code showing how I created the training file as an appendix. I created this code a while ago and have since added to it, so there are a bunch of other indicators and transforms that will be output to a csv file. You can pick out the ones that I’ve used in this post, or experiment with the others, as you like. The output file is called “variables. csv” and this is what I read into my training data object “gu_data” in this post.

Also, you will need to manually exclude some rows from the output file where indicator values are unable to be calculated due to an insufficient lookback period.

This code is written in Lite-C and is compatible with the Zorro platform, which is an awesome tool for doing this sort of analysis. Some exciting things you can do to extend my research by tweaking this code include:

1. Try other bar periods (I used daily)

2. If using daily bars, try sampling at a more sensible time (ie not midnight, as I have done here)

3. Obtain the same data for any market for which you have data by changing the asset() call towards the start of the script.

4. Use Zorro’s recently added pre-processing tools for better scaling/compressing/normalizing of the data than I have used here (these tools were added to the platform after I wrote this post)

5. Add other indicators or data that you are interested in.

6. Experiment with longer prediction horizons. By default, I used a one-day prediction horizon. Could predicting further into the future give better results?

7. Try using a classification approach by changing the target variable from a return (a numerical value) to a market direction (a binary value).

8. Plenty more that I haven’t even thought of…

Hope this is useful for you.

Ótimo trabalho! Very good tutorial for beginner like me!

A small question thought, why did you times 100 to the price change of the target?

Kris Longmore.

No reason other than aesthetics. I just find it nicer to look at figures on the left of the decimal point. You’ll get the same results if you don’t multiply by 100.

Thanks for reply.

Another small question bothers me while reading your post.

In “Removing highly correlated variables” step, we remained variables having correlation less than 0.5 with others. Then, why did some high correlation pairs still remain and show on the map? (f. i. velocity10 v. s. atrRat10 ; atrRat10 v. s. atrRat10_20)

Is there something possibly wrong?

Kris Longmore.

You can answer your own question very easily by looking into the documentation of the relevant function, caret::findCorrelation() . I’ll do the leg-work for you so that we have a reference here on Robot Wealth. From the documentation:

The absolute values of pair-wise correlations are considered. If two variables have a high correlation, the function looks at the mean absolute correlation of each variable and removes the variable with the largest mean absolute correlation.

The key point is that the function considers mean absolute correlations of each variable when the pair-wise correlation between them is higher than the threshold. The function will then remove the variable that has the highest mean absolute correlation with all remaining variables. This is not the same as simply removing any variable that has a pair-wise correlation with another variable greater than the cutoff value.

Espero que ajude.

I think the point is whenever two variables have a high correlation, one of them will be removed. The mean absolute correlations is just a way to decide which one would be cut out.

So, your correlation map shouldn’t have the pair like velocity10 to atrRat10 with large circle/high correlation. Otherwise, what’s the point for using the findCorrelation() and setting the cutoff ?

Kris Longmore.

The point is that the pair-wise correlations reveal useful information about the data. As do the results of the findCorrelation() function. Further, if there were many variables from which to select, the function is an efficient first step in processing them. Its also interesting to see which variables the function identifies for removal in relation to their pair-wise correlations with the remaining variables.

I don’t see any basis to your statement ‘your correlation map shouldn’t have the pair like velocity10 to atrRat10 with large circle/high correlation.’ Por que não? Does it not reveal useful information? Have I misrepresented the pair-wise correlations? I think you might be confusing the roles of the correlation map and the findCorrelation() function by assuming they represent the same thing, or perhaps you see them as being somehow contradictory rather than complimentary. I’m struggling to see where you’re coming from here.

I don’t think I’m confused with the function and the map, but still thanks for the reply. I’ll think it through again and check my code.

Vincenzo Somma.

I’m an R beginner and trying to reproduce the results.

I don’t understand where the “gu_data_filt” variable is filled with data in the first small code block of the article.

Qualquer ajuda foi apreciada.

Kris Longmore.

Vincenzo, you found an error – nicely spotted. There was a missing line of code which has been added.

I suggest adding “highCor <- highCor+1;" before "gu_data_filt <- gu_data[, - highCor];", for your reference.

Kris Longmore.

Can you explain that suggestion? Why do you need to increment the values in the highCor object?

It’s because we use cor(gu_data[, -1]), making highCor contained the wrong column numbers(all are 1 smaller) of removed feature. Adding 1 back so that we can get the correct result.

Kris Longmore.

Nicely spotted! You are absolutely correct. I’ve updated the code accordingly. Note that this would alter the features which get passed through to the next feature selection stage.

Vincenzo Somma.

If you have time can you update the article with the correct resulting features ?

Just to compare my results with yours.

Kris Longmore.

You mean the features that fall out of the correlation analysis? Or the entire post? Sure, I’ll update the figures and results, when I can find the time.

Note however that there aren’t any ‘correct’ features – a better way to approach the problem is to think in terms of which features or combination of features are most useful for the problem at hand.

Vincenzo Somma.

I mean in your reply to David pasted below, the resulting features are incorrect because of the code issue he outlined. Would be nice if you could update the article feature selection results after this fix.

Thanks for setting up a great interesting web site.

“Note that this would alter the features which get passed through to the next feature selection stage”

Kris Longmore.

Understood. Yes, updating the post is on my radar. Might take me a week or two to find the time, but I’ll make sure it happens.

Obrigado pelas palavras amáveis.

Kris Longmore.

In case you are following this thread via email, I’ve updated the post following David’s observations about the correlation filter.

Glad it helped. Looking forward to the updated result. Obrigado pelo bom trabalho.

Kris Longmore.

In case you are following this thread via email, I’ve updated the post following your observations about the correlation filter.

corburt erilio.

Ótimo post. I am facing a couple of these problems.

Hi Kris, thanks for releasing all the R code unified (including the GBM model), it’s very helpful and adds a lot of insight to the whole experimentation process.

Kris Longmore.

My pleasure! Glad to hear it is helpful.

Hi, thanks for the great article, but where is the link to download the scripts? In two places you say the scripts are available for download, but I can’t find any links.

Kris Longmore.

Você é muito bem-vindo. There’s a link at the top of the article, under Update 4. But you have to give me your email address in return. I promise not to spam you.

Predicting a single bar is bound to be problematic because of the aliasing noise. John Ehlers often points out that if you sample the data at one bar per day then the simple fact of sampling introduces noise that is significant for signal wavelengths less than 10 bars long.

I find the use of the normalize and zscore functions to be philosophically problematic for some sorts of data. As an example, lets take a value for the trend. Obviously if this value is positive, we have an up-trend. The problem is that normalize and zscore recenter the values around their mean, so a positive value might be transformed into a negative value. I’d be interested to see any studies on whether non-shifted values give better results.

Kris Longmore.

Its only bound to be problematic if you accept that Ehlers’ digital signal processing paradigm is applicable to financial time series. My personal opinion is that it can be, at times and under certain conditions. I’ve taken parts of Ehlers’ approach and used it in my own work, but I’d be careful of treating it as dogma. After all, many of the frequency decomposition techniques referenced in his work were intended for stationary, repeating signals, not the type of data we typically deal with. Of course they can be adapted to the markets, but a flexible and open minded approach is required.

Regarding the scaling and normalizing question: but the fact that such data, in its raw form, is of a certain sign is only applicable to our understanding and perception of what that means. A machine doesn’t care about the sign in the sense that we associate it with an ‘up’ tendência. It cares about what that data point means in relation to the other variables, in particular the target. But still, experimentation is the best way to explore such questions, and it wouldn’t be difficult to come up with a way to scale your data while preserving the sign. I’d love to hear what you find out.