How to use this page
Read the extract first, then the application and limits sections, and only then decide whether the thesis is strong enough for action or only for context.
Extractable overview
Feedback loop: prediction record -> market resolution -> Brier score and correctness -> weight updates.
Learning method: multiplicative weight updates can reward or punish analysts based on resolved outcomes.
Drift control: the learning layer includes safeguards that can pull a signal back toward neutral when its behavior changes.
What gets recorded
When EdgeVisor publishes a prediction, it can store the market id, category, estimate, market price, side, confidence, analyst confidences, signal metadata, and explanation payload. That creates a record that can later be reconciled with the actual market resolution.
This matters because learning without recorded state is theater. The system needs a stable memory of what it believed at prediction time.
| Recorded field | Why it matters |
|---|---|
| Estimate and market price | Define the exact disagreement the model acted on |
| Category and signal metadata | Allow later analysis of which market conditions helped or hurt the system |
| Explanation payload | Preserves what the product actually showed to the user, not just a hidden score |
How feedback updates weights
When a market resolves, the outcome tracker computes whether the thesis was correct and what the probability error looked like. Those signals can feed back into internal weighting. Stronger signals gain more influence when they are helpful; weaker ones lose influence when they mislead.
| Example | Predicted probability | Actual outcome | Brier score |
|---|---|---|---|
| Strong but wrong call | 0.80 | 0 | 0.64 |
| Measured correct call | 0.65 | 1 | 0.12 |
EdgeVisor also keeps regime summaries and rolling metrics so the learning loop is not just a single global score. In practice this means the system is trying to learn which mix of evidence is more useful under which market conditions instead of blindly trusting one internal pattern forever.
What the metrics do not mean
Brier score is not the same as profit. Win rate is not the same as calibration. And a good short-term patch does not guarantee that the same analyst mix will stay useful after the environment changes.
- High win rate can still be shallow: it may come from taking easy favorites rather than from well-calibrated probabilities.
- Good calibration can still lose money: execution timing and market microstructure still matter.
- Learning reduces blindness, not risk: it improves accountability, but it cannot remove liquidity risk, timing risk, or category mismatch.
Frequently asked questions
No. It can also use weaker feedback from resolved predictions, not just live execution outcomes.
No. Win rate and Brier score answer different questions. Calibration checks probability quality, not just directional correctness.