Direction argmax flip — the book flipped on zero new information

The predictor’s daily direction call flipped from 24 UP / 2 DOWN to 2 UP / 26 DOWN overnight, with no meaningful change in the underlying model scores. The bug wasn’t in a model — it was in how a probability got turned into a direction.

Date: 2026-07-06 · Severity: P1 · Detection-to-fix: same day

Symptoms

The daily direction tally — a standing sanity check on the predictor’s output — swung from a heavily UP day to a heavily DOWN day back to back, with the underlying calibrated scores barely moving. A near-total book reversal with no corresponding shift in signal is itself the anomaly; nothing about the market or the model had changed enough to justify it.

Detection

The direction counts are watched every morning as a standing distribution check, precisely because a healthy predictor’s day-to-day direction mix should be reasonably stable. A 24:2 → 2:26 swing was visible immediately in that count, before it reached any position-sizing decision downstream.

Root cause

Direction was derived by taking the argmax of a calibrated probability — effectively “UP if P(beat market) > 0.5, else DOWN.” That threshold implicitly assumes the label is balanced around 50/50. It isn’t: roughly 60% of names underperform SPY at the 21-day horizon the system targets, so the true zero-alpha crossing point sits away from p = 0.5. With the decision boundary planted at the wrong probability, small day-to-day recalibration noise near that boundary was enough to swing the argmax for most of the universe simultaneously — a book-wide reversal manufactured entirely by where the threshold was drawn, not by any change in the model’s actual view.

Fix

Direction is now derived directly from the sign of the predicted alpha itself — sign(predicted_alpha) — computed at a single derive_direction function and consumed at every site that needs a direction call. There is no longer a separate probability threshold to get wrong, and no FLAT state; the label is binary by construction, since “no opinion” was never actually a distinct trading thesis.

Systemic improvement

Collapsing direction derivation to one function, one rule, and one call site closes off this entire class of bug: any future change to calibration or thresholding can’t silently re-introduce a mismatched decision boundary, because there’s no longer a decision boundary to place — the model’s own sign is the answer.