Self-Improving AI Predictions

01 — THE LEARNING

Every wrong
prediction makes
the next one
better.

Unlike most forecasting tools, Polytragent remembers everything. prediction_store holds every past prediction with market context and outcome. When a market resolves, the background resolver records whether the AI was right or wrong. The AI sees this: "You predicted politics with 83% accuracy. You predicted crypto with 60%. You were overconfident on March predictions by 3%." It adjusts.

—

Full Accuracy History

Every prediction logged: market ID, AI forecast %, price at time of prediction, recommendation (BUY/HOLD), outcome. Overall win rate calculated continuously.

—

Calibration Bias Detection

AI detects systematic over/under confidence. "You're +3% overconfident on politics" or "-2% underconfident on economics". Future predictions are adjusted automatically.

—

Topic-Specific Accuracy

Win rate broken down by category: Politics (83%), Economics (85%), Crypto (60%), Geopolitics (71%). AI knows where it's strong and weak.

polytragent — accuracy review

polytragent — calibration adjustment

02 — THE IMPROVEMENT LOOP

The predict →
log → resolve →
learn → predict
cycle never stops.

This is the core Polytragent loop. Every prediction is logged. Every 6 hours, the background resolver checks Polymarket for newly resolved markets. Outcomes are matched to past predictions. Accuracy metrics are updated. Claude Sonnet sees the updated stats before the next /research command. The AI is literally learning from the trades it recommended.

Prediction Logged

User sends /research. AI makes 8–12 predictions. Each is logged to prediction_store with timestamp, market ID, AI estimate, Polymarket price, confidence level.

Wait for Resolution

The Polymarket event trades for days or weeks. Polytragent does nothing—unless the user asks for updates.

Background Resolver Runs

Every 6 hours, background.js hits the Polymarket API. Looks for newly resolved markets matching past predictions. Fetches YES/NO outcomes.

Match & Record Outcome

Compare resolved market ID to prediction_store. Mark CORRECT (AI said YES, market resolved YES) or WRONG. Record hypothetical P&L.

Update Calibration Metrics

Recalculate overall win rate, topic-specific accuracy, calibration biases. Flag improvements and regressions.

Claude Sees New Stats

Next /research command: Claude sees updated accuracy profile. If crypto weakness detected, confidence on crypto predictions is capped lower.

Improved Predictions

Recommendations are more calibrated, more accurate, more profitable. The loop reinforces itself.

How does
the AI learn?

How often does the background resolver check for resolutions?

Every 6 hours on a cron schedule. It's a non-blocking background job that checks Polymarket's API for newly resolved markets and matches them to past predictions.

What happens if a prediction is wrong?

The loss is recorded. Hypothetical P&L goes negative. The market and topic are flagged. On the next related prediction, confidence is reduced.

Can I see my prediction history?

Not yet. Future update will include a /history command showing all past predictions, outcomes, and P&L. This will be visible via Telegram dashboard.

Does the AI reset its learning?

No. prediction_store is permanent. The AI maintains a rolling 100-prediction window for recent performance + lifetime stats for long-term trends.

How much does accuracy improve over time?

Early users see ~5–10% improvement in calibration over 30 days of predictions. The first 20 predictions are the least accurate as the AI is still learning the user's market segment.

What if Polytragent has a bad streak?

Confidence levels are automatically reduced. A string of losses on "crypto" predictions will cap future crypto confidence at MEDIUM instead of HIGH, protecting against further losses.

Every wrongprediction makesthe next onebetter.