STOCK-RETURN-PREDICTION-USING-KNN-SVM-GUASSIAN-PROCESS-ADABOOST-TREE-REGRESSION-AND-QDA

Objective:

Predict stock stock price using Technical Indicators as predictors (features). Use Supervised Machine Learning Approach to predict stock prices. Employ the use of pipeline and GridSearch to select the best model Use Final Model to Predict Stock Returns. Show plots of stock Return Write Deployable script.

Note: That Every stock has different behaviour and so at every point we may have different best performing algorithm. For instance, after much testing Ranform Forest Algorithm perform better for predicting Apple Stocks than any other algo. Guassian process classifier performed better than every other algo at predicting IBM stocks etc.

Indicators/Predictors Used:

Moving Averages(Also called Rolling mean) Commodity Channel Index Momentum Stochastic Oscillator(D and K) Force Index Mass Index

# You can add ass many predictors are desired. # Most importantly if you have to do this, you may have to consider a feature selection using XGBoost. How to Use

>git clone https://github.com/kennedyCzar/STOCK-RETURN-PREDICTION-USING-KNN-SVM-GUASSIAN-PROCESS-ADABOOST-TREE-REGRESSION-AND-QDA Unpak the Files in a project folder Add File Path to Environment Variable using Spyder PythonPath Click on Synchronize with Environment. Restart Spyder. Report Issue

Output

plot of Feature Importance Gold Stock Retuns General Motors stock returns Apple stock returns Tesla Stock Returns

Estimation grid_RandomForestClassifier Best params: {'clfcriterion': 'gini', 'clfmax_depth': 8, 'clfminsamplesleaf': 8, 'clfminsamplessplit': 9} Best training accuracy: 0.855755894590846 Test set accuracy score for best params: 0.8546042003231018

Estimation gridRandomForestClassifierPCA Best params: {'clfcriterion': 'entropy', 'clfmax_depth': 7, 'clfminsamplesleaf': 6, 'clfminsamplessplit': 3} Best training accuracy: 0.7489597780859917 Test set accuracy score for best params: 0.691437802907916

Estimation grid_KNN Best params: {'clf_nneighbors': 10} Best training accuracy: 0.8037447988904299 Test set accuracy score for best params: 0.778675282714055

Estimation gridKNNPCA_ Best params: {'clf_nneighbors': 9} Best training accuracy: 0.7149791955617198 Test set accuracy score for best params: 0.6882067851373183

Estimation grid_SVC Best params: {'clfC': 5, 'clfgamma': 0.0001, 'clf__kernel': 'linear'} Best training accuracy: 0.8411927877947295 Test set accuracy score for best params: 0.851373182552504

Estimation gridSVCPCA Best params: {'clfC': 1, 'clfgamma': 1, 'clf__kernel': 'rbf'} Best training accuracy: 0.7323162274618585 Test set accuracy score for best params: 0.6865912762520194

Estimation grid_GaussianProcessClassifier Best params: {'clf_kernel': 1*2 RBF(lengthscale=1)} Best training accuracy: 0.8585298196948682 Test set accuracy score for best params: 0.8675282714054927

Estimation gridGaussianProcessClassifierPCA Best params: {'clf_kernel': 1*2 RBF(lengthscale=1)} Best training accuracy: 0.7295423023578363 Test set accuracy score for best params: 0.7011308562197092

Estimation grid_LogisticRegression Best params: {'clfC': 0.1, 'clfpenalty': 'l1', 'clf__solver': 'liblinear'} Best training accuracy: 0.8349514563106796 Test set accuracy score for best params: 0.8432956381260097

Estimation gridLogisticRegressionPCA Best params: {'clfC': 0.1, 'clfpenalty': 'l1', 'clf__solver': 'liblinear'} Best training accuracy: 0.7267683772538142 Test set accuracy score for best params: 0.7059773828756059

Estimation grid_DecisionTreeClassifier Best params: {'clf_maxdepth': 3} Best training accuracy: 0.8280166435506241 Test set accuracy score for best params: 0.8481421647819063

Estimation gridDecisionTreeClassifierPCA Best params: {'clf_maxdepth': 6} Best training accuracy: 0.7246879334257975 Test set accuracy score for best params: 0.6978998384491115

Estimation grid_AdaBoostClassifier Best params: {'clf_nestimators': 8} Best training accuracy: 0.8141470180305131 Test set accuracy score for best params: 0.8222940226171244

Estimation gridAdaBoostClassifierPCA Best params: {'clf_nestimators': 22} Best training accuracy: 0.6768377253814147 Test set accuracy score for best params: 0.6348949919224556

Estimation grid_GaussianNB Best params: {'clf__priors': None} Best training accuracy: 0.7441054091539528 Test set accuracy score for best params: 0.7544426494345718

Estimation gridGaussianNBPCA Best params: {'clf__priors': None} Best training accuracy: 0.7205270457697642 Test set accuracy score for best params: 0.7075928917609047

Estimation grid_QuadraticDiscriminantAnalysis Best params: {'clf__priors': None} Best training accuracy: 0.7933425797503467 Test set accuracy score for best params: 0.7883683360258481

Estimation gridQuadraticDiscriminantAnalysisPCA Best params: {'clf__priors': None} Best training accuracy: 0.7191400832177531 Test set accuracy score for best params: 0.7075928917609047

Classifier with best test set accuracy: grid_GaussianProcessClassifier

Conclusion

You must note that this strategy is trading is a low frequency approach and this 
fits to make steady income over a period of time.
For high Frequency Trading the result of the return is quite high.
GOLD happens to give the most return on applied strategy(As shown in the graphs above). Also worthy of mention is the fact that, Random Forest Classifier + PCA  in most cases performed better for stocks prices with both unsteady and steady rise.  Followed Next to Adaboost, then Gradientbost Classifier. In any case, the performance of an algorithm depends on the structure of  the underlying prices. Its behaviour over a time series. For different stocks different agorithm perform best.