Skip to content

Latest commit

 

History

History
120 lines (89 loc) · 5.65 KB

File metadata and controls

120 lines (89 loc) · 5.65 KB

Corrected winner-selection false-positive audit

Question

After correcting the historical price data, do false positives still materially damage the Ridge and XGBoost top-40 portfolios?

A false positive is defined here as a stock selected into the sector-capped top 40 whose realized five-session return is zero or negative. A severe false positive loses at least 10%.

Data-repair decomposition

The analysis separates:

  1. the published evaluation with contaminated labels;
  2. the original frozen scores with returns recomputed from corrected prices;
  3. fully refitted models using corrected prices, sectors, labels, and predictors.
Model and scenario Positive precision False-positive rate Severe false-positive rate Mean selected return Worst selected return
Ridge, published contaminated 53.94% 46.06% 3.45% 0.6375% -88.65%
Ridge, original scores with corrected returns 54.00% 46.00% 3.17% 0.4897% -33.58%
Ridge, corrected refit 53.65% 46.35% 3.53% 0.5395% -46.83%
XGBoost, published contaminated 52.79% 47.21% 3.82% 0.8335% -89.96%
XGBoost, original scores with corrected returns 52.77% 47.23% 3.53% 0.3828% -50.62%
XGBoost, corrected refit 51.59% 48.41% 3.93% 0.2216% -46.83%

The data error greatly exaggerated return magnitude, especially for XGBoost, but it did not create the false-positive rate. The rate is almost unchanged after corrected relabeling and remains high after complete refitting.

Relationship to the realized market

The table below classifies each holding window using the realized five-session SPY return and the realized return of the eligible universe. This is an attribution diagnostic, not information available when the portfolio is selected.

Model Realized market state Positive precision Severe false-positive rate Mean selected return Excess over universe
Ridge Broad market negative 29.12% 7.96% -2.58% -0.08%
Ridge Mixed 50.37% 1.69% -0.11% 0.00%
Ridge Broad market positive 74.22% 0.81% 3.25% 0.72%
XGBoost Broad market negative 26.85% 8.69% -2.86% -0.36%
XGBoost Mixed 47.72% 2.21% -0.24% -0.13%
XGBoost Broad market positive 72.56% 0.90% 2.82% 0.30%

False positives are strongly conditional on what the market subsequently does. This supports the economic intuition behind regime-aware risk control, but future market direction cannot itself be used as the regime input. The ranker also fails to preserve relative outperformance in broad negative markets, especially for XGBoost, so this is not merely an absolute-return labeling issue.

Decision-time regime

A causal regime uses only the decision-date SPY close and its trailing 200-session average. It enters weak below 98% of the average and returns to strong above 102%.

Model Decision-time regime Positive precision Severe false-positive rate Mean selected return
Ridge Strong 53.53% 1.98% 0.53%
Ridge Weak 53.93% 7.01% 0.57%
XGBoost Strong 51.65% 2.92% 0.29%
XGBoost Weak 51.47% 6.21% 0.08%

The causal regime does not improve ordinary precision. It does, however, identify a much higher severe-loss rate. Moving entirely to cash is not supported because Ridge still earns a positive average return in weak regimes. A narrower tail-risk guardrail is more consistent with the data.

Does a higher model score mean fewer false positives?

No stable safety gradient appears inside the selected 40.

Model and selected rank Positive precision Severe false-positive rate
Ridge, 1 to 10 55.00% 4.45%
Ridge, 31 to 40 54.40% 2.36%
XGBoost, 1 to 10 51.48% 5.16%
XGBoost, 31 to 40 50.71% 3.41%

The highest model scores are not the safest holdings. A rank-margin or score-confidence rule cannot be assumed to remove severe losses.

Is one stock responsible for portfolio losses?

Model Losing weeks Losses reversed by removing worst stock Median worst-stock share of weekly loss 90th percentile share
Ridge 84 7 18.33% 80.64%
XGBoost 81 6 14.91% 67.98%

A single stock is decisive in only about 7% to 8% of losing weeks. However, it can dominate the worst loss episodes. Severe false positives account for approximately 29.6% of Ridge's total negative holding-level drag and 29.4% of XGBoost's.

The corrected worst observations are genuine events, including the March 2020 collapse, EPAM in February 2022, Netflix in April 2022, and other event-driven losses. They are no longer split-scale artifacts.

Temporal stability

The weakest period is 2022 H1:

Model Positive precision Severe false-positive rate Mean selected return
Ridge 44.42% 5.87% -0.43%
XGBoost 42.12% 5.38% -0.89%

This supports requiring fold-level and year-level tail performance. Pooled precision alone hides important regime failures.

Decision

False positives remain a real model and market-regime problem after the data correction.

  • The prior false-positive concern is valid.
  • The corrupted data exaggerated losses but did not cause the high error rate.
  • XGBoost is worse than Ridge after correction.
  • Ordinary precision gating is not supported by the current probability model.
  • A causal weak-market signal is useful for identifying tail risk, not for predicting whether every selected stock will be positive.
  • The next model should target severe-loss recall under a bounded veto fraction.
  • Ridge top 40 remains the benchmark while that tail model is developed.