Time Series Analysis on Relative Value ETFs
In this Package we are using datascience / ML models to forecast relative returns for sector ETF's.
In particular we focus on Sector ETF performance relative to SPY (the broad market)
We have two notebooks and a folder structure to hold all our EDA, model testing/validation, & back test results:
- EDA_s0 holds our introductory scatter plots, shows how we compute all features*
- We show cointegration using ADF on relative returns
- Plot our features for visual confirmation/priming of modeling relationship
- Show how correlation can be used to split our ETF's into seperate datasets
- ModelFit_s2 holds our modeling process from start to finish
- This includes Random Forest/Logistic Regression Models
- Demo of using feature importance for financial markets research *
- Combinatorial Backtesting Method
- Walk Forward Backtesting Method
- The majority of our function calls reside in aFunctions.py
- Our API to gather Data, Side Tagging
- Cross Validation Functions for Feature Importance
- Updated Purged Method for MDA CV / SFI
- Plotting Functions
Sample Plots from our Analysis on Risk Off Tickers (['TLT' 'XLP' 'XLU' 'XLV']):
Mean Decrease in Importance (final iteration using Bagged Random Tree Model):
OOS = Out of Sample Accuracy, for MDA it is Purged and Cross Validated.
Mean Accuracy improvement over noised feature (MDA):
ROC for test set starting in 2017-06-30 - 8/2019:
IS Sharpe vs Out of Sample Sharpes for 3M time periods combinatorially trained, purged & measured using model predictions and test returns
Walk Forward Tabular Results:
We used 10 bp or 1/10th of a percent for transaction costs
| Walk forward Stats | Values |
|---|---|
| Ann-Return | 10.090 |
| Ann-Vol | 8.000 |
| Sharpe | 1.261 |
| Ann-Alpha-Bench | -0.540 |
| Ann-Alpha-Eq | 2.720 |
| TE-Bench | 8.807 |
| TE-EQwtd | 3.682 |
| IR-SPY | -0.061 |
| IR-EQ | 0.739 |
| TestYears | 4.670 |
| TCost | 10.000 |
- We used the ta package heavily for creation of technical features: TA
- We use Methods in our feature construction & modeling from Marcos Lopez de Prado Advances in Financial Machine Learning (2018)