Regime Filters Have Minimal Impact on Nearest Neighbor Cohorts
The article discusses the impact of adding 5 regime filters (VRP, VIX term structure, credit spread, yield curve, market breadth) to a cohort API. The results show the filters had little effect on the distribution of forward returns, suggesting the shape-based nearest neighbor search already captures similar regime information.
Why it matters
This article provides valuable insights for AI/ML agents that use historical market data and pattern matching. It highlights the limitations of simplistic regime filtering and the importance of understanding how the underlying data representation impacts the effectiveness of such techniques.
Key Points
- 15 regime filters had minimal impact on cohort distribution metrics like IQR and coverage
- 2Loose ±0.15 percentile bucketing retains 70% of the base pool, barely changing the top 200 neighbors
- 3Shape-based embeddings already capture similar regime information, making the filters redundant
Details
The article describes a test where 5 regime filters were added to a cohort API, with the goal of conditioning the forward return distributions. The filters were based on academic literature that suggests these factors (VRP, VIX term structure, credit spreads, yield curve, market breadth) should materially impact return predictability. However, the results showed the filters had very little effect - distribution widths, coverage, and cohort size barely changed from the baseline. The authors explain this is because the ±0.15 percentile bucketing used retains 70% of the base pool, and the shape-based nearest neighbor search already captures similar regime information. They suggest tighter bucketing (±0.05) or interaction filters (e.g. same_VRP AND same_term) may produce stronger effects, and regime-stratified calibration is likely where the real benefits will be seen.
No comments yet
Be the first to comment