Residual Analysis

Stacking models for better accuracy

If you are looking to compose two or more time series models for better accuracy, here are some resources. we_fix_bad_ml

Leaderboards

We've created leaderboards that might help you determine which models are useful in predicting the errors of other models. 

  • The TimeMachines residual leaderboards are produced using only z-streams. These might help inform your choices for point estimates.

  • The z1-streams (example) have their own leaderboards. Cross-reference with the overall earnings leaderboard to see if the code is public. 

The good thing is, adding a residual prediction to a point estimate can be accomplished in one line of code. 

Utilities

The TimeMachines package contains some utilities that allow you to tack on a residual model using one line of code. See skatertools/composition for the code that does this, though the usage example in skaters/simple/thinking.py might be easiest to follow and mimic. It is recommended that you read the timemachines README.md to grok the SKATER signature. 

 

New here?

We're trying to solve the last mile problem of time-series analysis.

Looking to write papers?

See Python module 1 for instructions on how to let your model roam a world of live time-series. 

Python Modules