Development Pipeline #ToDos

1. Get Data

2. Get Desired Pos

• Get desired pos change
• Get desired total pos
• Get a column of encoded values for classification models
  • For using predict_proba
  • using pd.qcut

3. Engineer Features

• Get all TALIB features
• Generate Time features
  • Is within open 10/15/20/30 minutes
  • Is within close 10/15/20/30 minutes
  • Is last day before holiday
  • Is first day after holiday
  • Is US holiday
• TODO other packages that generate features similarly

4. Select Features

• Recursive feature elimination using ExtraTreesClassifier
  • Using SelectFromModel

5. Train Model & Cross-Validation

• Use GridSearchCV for optimizing hyperparameters
• TODO: add DEAP functionality for faster computation
• Use TimeSeriesSplit in 5 parts. Also leaving .35 part out for final testing
• The model with the highest score is reported and saved.

6. Explain Model

• Use SHAP
• Use sklearn's plotting function

7. Pack & Ship for production