How good are our forecasts? Here's every number.
Two records, never mixed: the live record (forecasts written to an append-only ledger before the outcome, then scored automatically) and the research backtest (a simulated walk-forward test on history). Misses included.
Live record · DK-EWMA-LN v2
0 forecasts issuedLive record: every forecast is stored when issued and scored automatically when it matures. Nothing is edited or removed. Metrics appear once a horizon has at least 30 matured forecasts.
| Horizon | Evaluated | Pending | Effective n | 50% range held | 80% held | 90% held | Brier ±10% | Avg 80% width |
|---|---|---|---|---|---|---|---|---|
| 4 hours | 0 | 0 | 0 | Insufficient evidence for a reliable calibrated forecast — not enough matured forecasts yet. | ||||
| 24 hours | 0 | 0 | 0 | Insufficient evidence for a reliable calibrated forecast — not enough matured forecasts yet. | ||||
| 7 days | 0 | 0 | 0 | Insufficient evidence for a reliable calibrated forecast — not enough matured forecasts yet. | ||||
| 30 days | 0 | 0 | 0 | Insufficient evidence for a reliable calibrated forecast — not enough matured forecasts yet. | ||||
| 90 days | 0 | 0 | 0 | Insufficient evidence for a reliable calibrated forecast — not enough matured forecasts yet. | ||||
| 1 year | 0 | 0 | 0 | Insufficient evidence for a reliable calibrated forecast — not enough matured forecasts yet. | ||||
Directional hit rate is not reported: the model makes no up/down call, so there is nothing to score. Overlapping windows mean consecutive 7- and 30-day forecasts are correlated; "effective n" divides by the horizon length.
Research backtest (simulated)
walk-forward expanding window · warm-up 365 days · refit every 30 days · 10 assets (BTC, ETH, SOL, XRP, ADA, DOGE, LTC, LINK, AVAX, DOT) · data: Coinbase Exchange public candles · generated 2026-10-12
1-day horizon
| Model | n (effective) | Brier (up) | Skill vs coin flip | 95% CI | Calibration error | 50% / 80% / 90% held | Pinball |
|---|---|---|---|---|---|---|---|
| Naive baseline (coin flip + expanding-window volatility) | 20,620 (20,620) | 0.2500 | 0.00% | [0.000, 0.000] | 0.003 | 70.4% / 90.2% / 94.8% | 0.0101 |
| DK-EWMA-LN v1 | 20,620 (20,620) | 0.2500 | 0.00% | [0.000, 0.000] | 0.003 | 58.8% / 83.2% / 89.9% | 0.0095 |
| Momentum logistic regression | 20,620 (20,620) | 0.2519 | −0.77% | [-0.011, -0.005] | 0.031 | — | — |
| Historical simulation (trailing 365d) | 20,620 (20,620) | 0.2503 | −0.13% | [-0.004, 0.001] | 0.015 | 51.9% / 81.4% / 90.5% | 0.0096 |
| DK-EWMA-LN v2 | 20,620 (20,620) | 0.2500 | 0.00% | [0.000, 0.000] | 0.003 | 54.6% / 79.9% / 87.5% | 0.0095 |
| DK-EWMA-EQ v2 | 20,620 (20,620) | 0.2500 | 0.00% | [0.000, 0.000] | 0.003 | 50.3% / 79.9% / 89.9% | 0.0094 |
7-day horizon
| Model | n (effective) | Brier (up) | Skill vs coin flip | 95% CI | Calibration error | 50% / 80% / 90% held | Pinball |
|---|---|---|---|---|---|---|---|
| Naive baseline (coin flip + expanding-window volatility) | 20,560 (2,937) | 0.2500 | 0.00% | [0.000, 0.000] | 0.011 | 67.3% / 89.9% / 94.8% | 0.0266 |
| DK-EWMA-LN v1 | 20,560 (2,937) | 0.2500 | 0.00% | [0.000, 0.000] | 0.011 | 55.5% / 80.6% / 88.1% | 0.0260 |
| Momentum logistic regression | 20,560 (2,937) | 0.2596 | −3.85% | [-0.053, -0.026] | 0.060 | — | — |
| Historical simulation (trailing 365d) | 20,560 (2,937) | 0.2524 | −0.95% | [-0.022, 0.002] | 0.040 | 51.9% / 80.3% / 89.5% | 0.0259 |
| DK-EWMA-LN v2 | 20,560 (2,937) | 0.2500 | 0.00% | [0.000, 0.000] | 0.011 | 56.0% / 80.7% / 88.3% | 0.0261 |
| DK-EWMA-EQ v2 | 20,560 (2,937) | 0.2500 | 0.00% | [0.000, 0.000] | 0.011 | 50.8% / 80.7% / 90.6% | 0.0261 |
30-day horizon
| Model | n (effective) | Brier (up) | Skill vs coin flip | 95% CI | Calibration error | 50% / 80% / 90% held | Pinball |
|---|---|---|---|---|---|---|---|
| Naive baseline (coin flip + expanding-window volatility) | 20,330 (678) | 0.2500 | 0.00% | [0.000, 0.000] | 0.019 | 63.3% / 88.5% / 94.8% | 0.0566 |
| DK-EWMA-LN v1 | 20,330 (678) | 0.2500 | 0.00% | [0.000, 0.000] | 0.019 | 51.9% / 76.3% / 84.2% | 0.0575 |
| Momentum logistic regression | 20,330 (678) | 0.2774 | −10.96% | [-0.157, -0.061] | 0.130 | — | — |
| Historical simulation (trailing 365d) | 20,330 (678) | 0.2652 | −6.06% | [-0.113, 0.001] | 0.099 | 48.6% / 75.9% / 84.9% | 0.0596 |
| DK-EWMA-LN v2 | 20,330 (678) | 0.2500 | 0.00% | [0.000, 0.000] | 0.019 | 58.8% / 81.8% / 88.8% | 0.0590 |
| DK-EWMA-EQ v2 | 20,330 (678) | 0.2500 | 0.00% | [0.000, 0.000] | 0.019 | 54.2% / 81.8% / 91.1% | 0.0589 |
90-day horizon
| Model | n (effective) | Brier (up) | Skill vs coin flip | 95% CI | Calibration error | 50% / 80% / 90% held | Pinball |
|---|---|---|---|---|---|---|---|
| Naive baseline (coin flip + expanding-window volatility) | 19,730 (219) | 0.2500 | 0.00% | [0.000, 0.000] | 0.043 | 58.6% / 88.7% / 94.7% | 0.1002 |
| DK-EWMA-LN v1 | 19,730 (219) | 0.2500 | 0.00% | [0.000, 0.000] | 0.043 | 46.1% / 75.3% / 85.3% | 0.1024 |
| Momentum logistic regression | 19,730 (219) | 0.3206 | −28.24% | [-0.424, -0.143] | 0.231 | — | — |
| Historical simulation (trailing 365d) | 19,730 (219) | 0.3158 | −26.33% | [-0.404, -0.079] | 0.212 | 42.1% / 64.0% / 71.2% | 0.1275 |
| DK-EWMA-LN v2 | 19,730 (219) | 0.2500 | 0.00% | [0.000, 0.000] | 0.043 | 55.6% / 84.3% / 90.9% | 0.1068 |
| DK-EWMA-EQ v2 | 19,730 (219) | 0.2500 | 0.00% | [0.000, 0.000] | 0.043 | 50.9% / 84.2% / 91.9% | 0.1084 |
365-day horizon
| Model | n (effective) | Brier (up) | Skill vs coin flip | 95% CI | Calibration error | 50% / 80% / 90% held | Pinball |
|---|---|---|---|---|---|---|---|
| Naive baseline (coin flip + expanding-window volatility) | 16,980 (47) | 0.2500 | 0.00% | [0.000, 0.000] | 0.013 | 60.4% / 88.4% / 94.2% | 0.1984 |
| DK-EWMA-LN v1 | 16,980 (47) | 0.2500 | 0.00% | [0.000, 0.000] | 0.013 | 46.1% / 76.3% / 85.5% | 0.1983 |
| DK-EWMA-LN v2 | 16,980 (47) | 0.2500 | 0.00% | [0.000, 0.000] | 0.013 | 56.0% / 82.6% / 88.8% | 0.2100 |
| DK-EWMA-EQ v2 | 16,980 (47) | 0.2500 | 0.00% | [0.000, 0.000] | 0.013 | 56.7% / 82.6% / 89.3% | 0.2132 |
Verdicts (computed by rule, not edited)
- DK-EWMA-LN v1 · 1d — ranges: calibrated direction: no edgeNo directional view by construction (P(up)=0.5), so Brier skill is exactly 0. Interval coverage 50/80/90%: 58.8%/83.2%/89.9% (within the +/-5pp tolerance at 80% and 90%).
- Momentum logistic regression · 1d — ranges: n/a direction: no edgePooled Brier skill -0.0077, 95% block-bootstrap CI [-0.0108, -0.0046] lies below zero: no directional edge over a coin flip; the model is worse than the baseline. No interval forecast; calibration not applicable.
- DK-EWMA-LN v1 · 7d — ranges: calibrated direction: no edgeNo directional view by construction (P(up)=0.5), so Brier skill is exactly 0. Interval coverage 50/80/90%: 55.5%/80.6%/88.1% (within the +/-5pp tolerance at 80% and 90%). Overlapping 7-day windows: effective sample ~2937 (fewer still given cross-asset correlation).
- Momentum logistic regression · 7d — ranges: n/a direction: no edgePooled Brier skill -0.0385, 95% block-bootstrap CI [-0.0525, -0.0255] lies below zero: no directional edge over a coin flip; the model is worse than the baseline. No interval forecast; calibration not applicable. Overlapping 7-day windows: effective sample ~2937 (fewer still given cross-asset correlation).
- DK-EWMA-LN v1 · 30d — ranges: not calibrated direction: no edgeNo directional view by construction (P(up)=0.5), so Brier skill is exactly 0. Interval coverage 50/80/90%: 51.9%/76.3%/84.2% (outside the +/-5pp tolerance at 80% and 90%). Intervals are too narrow (under-coverage, fat tails). Overlapping 30-day windows: effective sample ~678 (fewer still given cross-asset correlation).
- Momentum logistic regression · 30d — ranges: n/a direction: no edgePooled Brier skill -0.1096, 95% block-bootstrap CI [-0.1573, -0.0614] lies below zero: no directional edge over a coin flip; the model is worse than the baseline. No interval forecast; calibration not applicable. Overlapping 30-day windows: effective sample ~678 (fewer still given cross-asset correlation).
- DK-EWMA-LN v1 · 90d — ranges: calibrated direction: no edgeNo directional view by construction (P(up)=0.5), so Brier skill is exactly 0. Interval coverage 50/80/90%: 46.1%/75.3%/85.3% (within the +/-5pp tolerance at 80% and 90%). Overlapping 90-day windows: effective sample ~219 (fewer still given cross-asset correlation).
- Momentum logistic regression · 90d — ranges: n/a direction: no edgePooled Brier skill -0.2824, 95% block-bootstrap CI [-0.4243, -0.1430] lies below zero: no directional edge over a coin flip; the model is worse than the baseline. No interval forecast; calibration not applicable. Overlapping 90-day windows: effective sample ~219 (fewer still given cross-asset correlation).
- DK-EWMA-LN v1 · 365d — ranges: n/a direction: no edgeReported, not verdicted: only ~7 non-overlapping 365-day windows in the sample (summed per-asset n_eff 47, but the assets move together), below the 20 needed for a verdict. Coverage 50/80/90%: 46.1%/76.2%/85.5%. Show only as an uncalibrated scenario.
Known biases
The backtest uses 10 assets still listed today (survivorship bias), overlapping windows, and one exchange's prices. It is evidence, not proof. The live record above is the real test. Reproduce it: python3 cli.py fetch && python3 cli.py backtest --out reports/backtest.json