Learning Loop

Autonomous model improvement — track predictions, measure outcomes, recalibrate

Prediction Snapshots
0
No snapshots yet
Outcome Records
0
0 collected
Metrics Computed
0
Awaiting data
Calibration Version
v2
✓ Stable
Current Calibration Weightsv2

conservative Profile

pe
9.5%
growth
14.2%
margin
10.3%
debt
7.9%
quality
23.7%
valuation
13.4%
weinstein
12.1%
sentiment
2.5%
analyst
6.4%

aggressive Profile

pe
9.5%
growth
25.3%
margin
6.3%
debt
4.7%
quality
14.2%
valuation
19.0%
weinstein
12.1%
sentiment
2.5%
analyst
6.4%
■ Scoring levers (Weinstein, Sentiment, Analyst) — derived from 30-year ElasticNet backtest, IC-weighted across 3 horizons
Drift Detection
Model Stable

Insufficient metrics data (0 records, need at least 2)

Model Performance Metrics

No metrics yet — take a snapshot and wait for outcomes to mature

First metrics will appear after 3-month outcomes are collected

Calibration History

No calibration changes yet

The initial calibration (v2) was derived from the 30-year ElasticNet backtest

How the Learning Loop Works
1

Snapshot

Weekly capture of all stock scores, lever values, and current prices. Creates a timestamped prediction record.

2

Wait & Collect

At 3m, 6m, and 1y after each snapshot, actual prices are fetched and returns computed against predictions.

3

Measure

IC (rank correlation), hit rate, and long-short spread are computed per profile and horizon. Per-lever ICs identify which factors are working.

4

Recalibrate

When drift exceeds 15%, lever weights are adjusted proportionally to their IC contribution. Changes are logged with full justification.