Share a powerful tool for factor Mining: genetic programming

How do you mine factors ？ Based on experience ？ But experience is limited , There will always be a useful time . Based on public materials such as research reports or papers ？ But this kind of factor inevitably involves factor congestion , After all, effective factors , Other people can also use .

 

So is there any other way ？ The answer is yes. .

 

Today we are based on the 《 Stock selection factor mining based on genetic programming in artificial intelligence series 》, Let's introduce a A sharp tool for factor mining ：  Genetic programming  .

 

What is genetic programming ？

 

Genetic programming is a branch of evolutionary algorithm , It is a heuristic formula evolution technique . It starts with a random group of formulas . By simulating the process of genetic evolution in nature , To gradually generate formula groups that fit specific goals . As a supervised learning method , Genetic programming can be based on specific goals , Find some hidden 、 A mathematical formula that is difficult to construct through the human brain . The traditional supervised learning algorithm is mainly used to fit the relationship between features and tags , Genetic programming is more applied to feature mining （ Feature Engineering ）.

——《 Stock selection factor mining based on genetic programming in artificial intelligence series analysis report 》

 

Previous factor studies have been “ First there is logic , Then there is the formula ”, It's a kind of “ Deductive method ”. But the form of genetic programming is “ First there is the formula , Then there is logic ”, Belong to “ Induction ”. Its advantage is that it can make full use of the powerful computing power of the computer to carry out heuristic search , At the same time, it breaks through the limitations of human thinking , Dig out some hidden 、 Factors that are difficult to construct through the human brain , Provide more possibilities for factor research .

 

Genetic evolution in organisms involves the inheritance of genes , variation , Adaptability to the ecological environment, etc , The same is true in genetic programming algorithms , There will also be cross variation 、 Subtree variation 、 A little variation 、Hoist Variation and fitness, etc , For specific details, please refer to the research report or thesis .

 

We use Python In the genetic programming project gplearn Module package for factor mining , The main parameters of the model are as follows ：

 

 

The data used in the model are as follows ：

 

• Test varieties ： The Shanghai composite index
• Back test interval ：2010 year 01 month 01 Japan -2022 year 05 month 31 Japan
• The initial factor ： Opening price 、 Closing price 、 Highest price 、 The lowest price 、 volume 、 Yield 、 Volume weighted average price
• Predict the goal ： future 5 Day yield
• Function list ： all gplearn Built-in function

 

Once the data is ready, you can start training the model ：

 

gp1 = SymbolicTransformer(generations=10, population_size=1000, function_set=function_set, init_depth=(1,4), tournament_size=20, metric='spearman', p_crossover=0.4, 
                          p_subtree_mutation=0.01, p_hoist_mutation=0, p_point_mutation=0.01, p_point_replace=0.40, 
                          warm_start=False, verbose=1,random_state=0, n_jobs=-1,feature_names=['open', 'close', 'high', 'low', 'volume', 'return_rate', 'vwap'])
...
gp1.fit(train,label)#  Training models 

 

The model will automatically display the process log , among Fitness It's fitness , What we choose here is Spearman Rank correlation coefficient , The higher the correlation coefficient , Represents factors and the future 5 The higher the correlation of daily yield  .

 

  

We further show the iterative process of the optimal factor in the form of a curve ：

 

 

As can be seen from the figure above , The optimal factor is approximately iterated to the fourth generation （X In the shaft ,0 It's the first generation ） When , The rank correlation coefficient reached a higher level , Subsequent iterations do not improve much .

  

Finally, through the tree diagram, we can see the optimal factor of the model iteration ：

  

  

To express it by formula is ：log( Closing price )/log( volume ) . Combine the top ten optimal factors of the model ：

 

   

You can find , There are many repeated factors in the output of the model , After removing the repetition factor , Only two factors are ：log( Closing price )/log( volume ) and log( volume )/log( Closing price ) .

 

In fact, these two factors should be the same factor , Just a reciprocal deformation . With log( Closing price )/log( volume ) Factor view , First, calculate the logarithm of closing price and trading volume respectively , To divide again , It can be regarded as the closing price weighted by the reciprocal of trading volume . Interested friends , The performance of this factor can be further tested , Factor mining can also be carried out for other indexes or commodity futures .

 

This article is a preliminary exploration of genetic programming , But there is still a big piece of content that has not been solved yet , For example, the functions used this time are gplearn Built-in function . How to extend functions ？ Especially the time series function , For example, seeking history 5 Daily mean . A single variety of the current test variety , How to extend to multiple varieties ？ How to deal with such 3D data ？ All these need to be solved .

 

Later, an advanced version of genetic programming will be launched , Take you further to explore factor mining ！