Overfitting in Return-Pattern Strategies and EWMAC Forecast Scalars
Summary
The post illustrates overfitting with a strategy that classifies recent price movements into patterns and estimates the following month's average return for each pattern. It progressively divides a 64-business-day lookback into more segments, creating more possible states. A long-only version buys when the current state historically exceeded the average return and otherwise remains flat. The Microsoft example compares a model fitted on all available history with one fitted only through 2015 and then tested afterward, as well as applying the Microsoft-fitted method to Apple.
The highly segmented approach looks successful in-sample but deteriorates on later unseen data and transfers poorly to another stock, which the author attributes to fitting the historical details too closely. The post then introduces a simpler 64-day model based on whether prices rose or fell, but the excerpt ends before showing its results. It also ends at a heading about calculating EWMAC forecast scalars without providing that method. Thus, the excerpt demonstrates a useful train/test and cross-instrument caution, but leaves its simpler-model comparison and scalar discussion unfinished.
Key ideas
- Dividing a lookback period into more segments increases the number of return patterns a model can fit.
- The example trades long when a pattern's historical forward return exceeds the average and otherwise stays flat.
- A finely segmented Microsoft model performs poorly after its fitting period and when applied to Apple.
- Out-of-sample and cross-instrument results expose weaknesses hidden by fitting on the full history.
- The excerpt introduces a simpler trend-based model and EWMAC scalar topic without reporting their results.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.