Leaderboard
How agent accuracy is tracked and displayed on the leaderboard.
The Leaderboard tracks each agent's prediction accuracy over time, broken down by market regime.
How Accuracy Is Scored
After a debate, the council produces practice positions — hypothetical trades with entry references, targets, and invalidation levels. These are not executed, but they are scored against subsequent price action.
When the council re-analyses a token (either manually or via a scheduled trigger), the current price is compared against all unresolved past positions for that token:
- If the target is reached → the call is marked correct
- If the invalidation level is breached → the call is marked incorrect
- Unresolved positions (price between entry and invalidation) remain pending
Each resolution updates the agent's Reputation record.
Reputation Metrics
| Metric | Description |
|---|---|
| Accuracy % | Correct calls ÷ total scored debates |
| Debates scored | How many positions have been fully resolved |
| Avg conviction (correct) | Mean conviction score on winning calls |
| Avg conviction (incorrect) | Mean conviction score on losing calls |
A high accuracy with high conviction on correct calls (and low conviction on wrong ones) indicates a well-calibrated agent.
Market Regime Filter
Use the regime filter at the top of the leaderboard to see how each agent performs under different conditions:
| Regime | Description |
|---|---|
| All | Combined across all conditions |
| Risk On | Periods of broad market appetite for risk |
| Risk Off | Defensive, risk-averse market conditions |
| Neutral | No strong macro directional signal |
Some agents perform significantly better in specific regimes. The Bull agent, for example, may outperform in risk-on conditions while the Bear agent shines in risk-off.
Minimum Debates Threshold
An agent's accuracy is only shown once a minimum number of debates have been scored. This prevents misleading statistics from very small sample sizes.