Model Accuracy — Audited Performance
Every Kronos prediction is stored and later scored against actual prices. The /accuracy endpoint returns audited performance metrics — hit rate, Brier score, calibration, and baselines.
Metrics
| Metric | Description |
|---|---|
| Hit rate | Percentage of correct direction predictions |
| Brier score | Mean squared error of probabilistic predictions (lower is better) |
| Calibration | How well predicted probabilities match observed frequencies |
| Coverage | Percentage of time periods with active predictions |
| Edge vs no-change | Performance vs a naive "no change" baseline |
| Edge vs momentum | Performance vs a naive momentum baseline |
- Hit rate
- The percentage of predictions where the direction was correct. A 58% hit rate means 58 out of 100 predictions called the right direction.
- Brier score
- A measure of how accurate probabilistic predictions are. Lower is better. A Brier score of 0.21 is good; 0.5 is random chance for binary outcomes.
- Calibration
- How well predicted probabilities match observed frequencies. If the model says "70% up" and 70% of those predictions go up, calibration is perfect.
- Conformal prediction
- A statistical framework that produces prediction intervals with guaranteed coverage. Kronos uses it to generate p05-p95 price ranges.
API access
# Model accuracy (paid — $0.001)
curl -H "X-PAYMENT: ..." \
https://kronos.seshat.markets/api/feeds/kronos/accuracy?symbol=btc_usdt
# Per-timeframe candle audit (paid — $0.005)
curl -H "X-PAYMENT: ..." \
https://kronos.seshat.markets/api/feeds/kronos/accuracy/candles?symbol=btc_usdt
# Free preview (no payment)
curl https://kronos.seshat.markets/api/feeds/kronos/accuracy-preview/btc_usdtHow are predictions scored against actual prices?
Each prediction is stored with a timestamp and horizon. After the horizon passes, the actual price is fetched from the exchange. The prediction is scored for direction accuracy, Brier score, and whether the outcome fell within the conformal range.
- Kronos generates a forecast at time
Twith horizonH - The prediction is stored with its conformal ranges and probabilities
- After
Hhours, the actual price is fetched from the exchange - The prediction is scored: direction correct? Brier score? Within conformal range?
- Results are aggregated into the accuracy metrics
All scoring is automated and auditable. Every prediction has a decision_id that can be traced from forecast to resolution.
What is the risk state and why does it matter?
Kronos tracks a risk state based on recent prediction streaks: Normal (full confidence), Cooldown (reduced confidence after losses), or Suspended (predictions paused). It protects users from a degrading model by automatically reducing output confidence.
- Normal — operating normally, full confidence multiplier
- Cooldown — recent losses, reduced confidence multiplier
- Suspended — extended losses, predictions paused
Check the current risk state: GET /api/feeds/kronos/risk (free)
Use cases
Model Trust Calibration
Before following any signal, check the hit rate and Brier score. A 58% hit rate with 0.21 Brier means the model is calibrated and reliable. A 52% hit rate with 0.35 Brier means dial down trust.
Edge Verification
Compare edge_vs_no_change and edge_vs_momentum. If the model beats both baselines, it has genuine predictive edge. If it only beats no-change but loses to momentum, the model adds little over a simple trend-following strategy.
Per-Symbol Performance
Filter by ?symbol=btc_usdt. If BTC has 64% accuracy but SOL has 48%, only trade BTC signals. Use per-symbol accuracy to decide which assets to follow and which to ignore.
Calibration-Guided Sizing
If calibration status is "good", the model's confidence estimates are accurate — size positions by confidence. If calibration is "poor", confidence is unreliable — use fixed sizing instead.
Frequently asked questions
How accurate is the Kronos crypto prediction model?
What is a Brier score for crypto predictions?
How do I verify crypto prediction model accuracy?
GET /api/feeds/kronos/accuracy?symbol=btc_usdt ($0.001) for audited hit rate, Brier score, and calibration. A free preview is available at /api/feeds/kronos/accuracy-preview/btc_usdt. Every prediction has a traceable decision_id.Does the crypto model beat naive baselines?
edge_vs_no_change and edge_vs_momentum. If the model beats both baselines, it has genuine predictive edge. If it only beats no-change but loses to momentum, the model adds little over a simple trend-following strategy.Can I check per-symbol crypto prediction accuracy?
?symbol=btc_usdt. If BTC has 64% accuracy but SOL has 48%, only trade BTC signals. Use per-symbol accuracy to decide which assets to follow and which to ignore. See /accuracy-candles for per-timeframe MAPE.