Predictor meta-model collapse — 27 UP / 0 DOWN

A weekly retrain produced a meta-learner with a degenerate output distribution. Catching it before any trading decisions used the bad model — and finding the structural cause rather than symptom-patching — illustrates the kind of “loud failure beats silent drift” discipline this system is built for.

Date: 2026-04-28 · Severity: P1 · Resolution: 5-PR arc, same week

Symptoms

Morning predictor email showed 27 UP predictions and 0 DOWN predictions across the universe. Confidence values clustered into a narrow band with only a few unique values. A healthy distribution looks roughly balanced with continuous confidence values across the full range.

Detection

The morning briefing email itself was the detector — the distribution shape was sharply different from the typical run. Caught before the executor used the predictions for sizing or veto decisions, so no trading impact.

Root cause

The Layer-2 Ridge meta-learner combines outputs from Layer-1 specialized models with research-context features (research composite score, conviction signal, sector macro modifier) sourced from the research module’s signals.json. The research module was emitting these values correctly — the bug was on the consumption side. During an earlier scaffold pass inside the predictor’s meta-trainer, these features had been hardcoded to placeholder constant values in the walk-forward training loop, with the intent to swap in the real per-ticker reads from signals.json once that data path was finalized. The data path landed but the swap was missed.

With constant inputs, Ridge correctly assigned zero coefficients to those features — a healthy regularizer doing exactly what it was supposed to do. The meta-learner then effectively reduced to its remaining Layer-1 inputs, producing a much narrower decision surface than it was designed to.

Fix

Five-PR arc, all on alpha-engine-predictor — no changes to the research module or to signals.json:

• Replace the hardcoded constants in meta_trainer.py with real per-ticker reads from signals.json for the walk-forward training loop
• Streaming refactor for memory efficiency (an OOM issue surfaced as a side-effect during the fix; closed in the same arc rather than papered over)
• Three retraining attempts, each gated on a named-baseline check before promotion

The final promoted model showed the research-context coefficients non-zero — the structural diagnostic that the upstream wiring was now load-bearing.

Systemic improvement

Validation IC moved from 0.053 to 0.132 — a 2.48× improvement on the predictive substrate the executor relies on. Beyond the immediate fix, the arc surfaced a class of bug — placeholder constants in production training data — that drove a follow-up audit pass for similar dead-input patterns elsewhere in the training pipelines.