Backtest results
An honest replay of the direction model on past data. For every coin and day in our history, we run the model and check what price actually did 1, 3, or 7 days later— only on days the model hadn't learned from. Pick a coin group and horizon to see where it has real edge and which stress / pulse conditions give the most reliable calls.
Meaningful = predictions where the model's probability was at least ±3% away from 50/50 — i.e., the model actually picked a side. The all-calls number includes every prediction, even near-coinflip ones.
Accuracy by Stress band
Each row groups predictions by the pair's stress score on that day. Look for bands where meaningful accuracy is well above 50% — those are the regimes where the model actually has an edge.
| Stress band | Calls | All-calls acc | Meaningful acc | Edge |
|---|---|---|---|---|
| 0.00-0.25 | 41(25 meaningful) | 65.9% | 64.0% | +0.0pp |
| 0.25-0.50 | 1,972(1,381 meaningful) | 53.4% | 53.5% | +0.1pp |
| 0.50-0.75 | 1,590(1,286 meaningful) | 57.4% | 57.6% | -0.1pp |
| 0.75-1.01 | 454(415 meaningful) | 65.0% | 64.3% | +0.0pp |
Accuracy by Pulse band
Same idea, grouped by the pulse score (how fast positioning, funding and volatility are shifting versus normal). Pulse tends to lead the move — does the model trust fast-moving setups more?
| Pulse band | Calls | All-calls acc | Meaningful acc | Edge |
|---|---|---|---|---|
| 0.00-0.25 | 68(49 meaningful) | 55.9% | 53.1% | +0.0pp |
| 0.25-0.50 | 1,367(983 meaningful) | 57.4% | 57.3% | +0.2pp |
| 0.50-0.75 | 2,153(1,698 meaningful) | 55.4% | 55.9% | -0.1pp |
| 0.75-1.01 | 469(377 meaningful) | 58.2% | 59.4% | +0.0pp |
Stress × Pulse accuracy heatmap
Each cell is the model's hit rate when stress (rows) and pulse (cols) landed in those bands together, with the number of calls (n) underneath. Green = real skill (beat the base rate); gray = the hit rate is just the base rate, not skill. Hover any cell for the baseline and confidence interval.
| Pulse → | ||||
|---|---|---|---|---|
| Stress ↓ | 0.00-0.25 | 0.25-0.50 | 0.50-0.75 | 0.75-1.01 |
| 0.00-0.25 | — n=0 | 59% n=17 | 75% n=8 | — n=0 |
| 0.25-0.50 | 48% n=21 | 53% n=506 | 54% n=743 | 54% n=111 |
| 0.50-0.75 | 52% n=23 | 62% n=374 | 55% n=700 | 58% n=189 |
| 0.75-1.01 | 80% n=5 | 62% n=86 | 63% n=247 | 70% n=77 |
Cross-sectional model — “will it beat the median coin today?”
Trained on relative strength versus the same-day median of all coins, so the market's overall drift is removed from the label by construction — the baseline is exactly 50%, and anything above it is pure per-coin skill. Honest walk-forward numbers, meaningful calls only.
Breakout model — “will it break ±1.5σ within N days?”
Predicts whether a coin will touch a volatility-scaled barrier (±1.5× its own daily volatility) within the window. Globally it only matches the base rate — so the product shows it only for pairs in the conditions below, where the backtest shows real edge over the event base rate.
Predictions that worked — the model's lean matched what price did
Concrete examples where the model called ▲ up and price went up, or called ▼ down and price went down — within the chosen horizon. Ordered by realized move size, so the impressive calls float to the top. Pure 50/50 picks are excluded (|p − 50%| ≥ 3%).
| Date | Pair | Lean | Realized |
|---|---|---|---|
| 2026-04-06 | RAVEUSDT | 54% | +3143.21% |
| 2026-04-05 | RAVEUSDT | 55% | +2271.33% |
| 2026-01-24 | BULLAUSDT | 59% | +938.27% |
| 2026-04-04 | RAVEUSDT | 55% | +770.81% |
| 2026-04-29 | LABUSDT | 53% | +580.13% |
| 2026-04-30 | LABUSDT | 55% | +546.06% |
| 2026-04-03 | RAVEUSDT | 54% | +526.61% |
| 2026-06-04 | BEATUSDT | 55% | +487.91% |
| 2026-06-04 | ESPORTSUSDT | 57% | +451.90% |
| 2026-06-03 | BEATUSDT | 58% | +434.50% |
| 2026-04-06 | BLESSUSDT | 54% | +417.38% |
| 2026-01-13 | FHEUSDT | 54% | +415.23% |
| 2026-04-03 | SIRENUSDT | 54% | +400.69% |
| 2026-06-05 | BEATUSDT | 55% | +388.93% |
| 2026-01-12 | FHEUSDT | 55% | +355.10% |
| 2026-02-19 | POWERUSDT | 55% | +354.62% |
| 2026-04-27 | SKYAIUSDT | 54% | +336.94% |
| 2026-06-02 | VELVETUSDT | 54% | +333.75% |
| 2026-06-02 | BEATUSDT | 60% | +333.43% |
| 2026-03-28 | DUSDT | 55% | +299.05% |
| 2026-05-24 | PORTALUSDT | 56% | +295.01% |
| 2026-01-23 | BULLAUSDT | 57% | +294.97% |
| 2026-04-28 | LABUSDT | 54% | +290.68% |
| 2026-06-01 | BEATUSDT | 56% | +269.12% |
| 2026-05-25 | LABUSDT | 55% | +265.24% |
| 2026-04-08 | AKEUSDT | 55% | +259.77% |
| 2026-04-09 | ORDIUSDT | 56% | +252.02% |
| 2026-04-07 | BLESSUSDT | 54% | +251.25% |
| 2026-05-26 | LABUSDT | 53% | +245.96% |
| 2026-06-01 | VELVETUSDT | 56% | +244.97% |
| 2026-05-27 | LABUSDT | 53% | +243.59% |
| 2026-01-10 | FHEUSDT | 55% | +240.39% |
| 2026-04-18 | BSBUSDT | 54% | +237.20% |
| 2026-01-31 | SIRENUSDT | 53% | +229.93% |
| 2025-12-30 | BROCCOLI714USDT | 55% | +221.98% |
| 2026-04-25 | UBUSDT | 55% | +218.53% |
| 2026-02-05 | PIPPINUSDT | 53% | +215.36% |
| 2026-04-24 | BUSDT | 55% | +213.17% |
| 2026-01-22 | BULLAUSDT | 56% | +211.23% |
| 2026-05-26 | PORTALUSDT | 55% | +210.49% |
| 2026-05-24 | ALLOUSDT | 54% | +208.87% |
| 2026-05-14 | EDENUSDT | 56% | +208.76% |
| 2026-06-05 | ESPORTSUSDT | 56% | +208.12% |
| 2026-05-15 | EDENUSDT | 56% | +204.93% |
| 2026-03-27 | AIOTUSDT | 56% | +201.37% |
| 2026-04-10 | ORDIUSDT | 56% | +192.54% |
| 2025-12-04 | FHEUSDT | 54% | +192.12% |
| 2026-04-22 | AIOTUSDT | 55% | +188.42% |
| 2026-03-19 | BRUSDT | 54% | +187.75% |
| 2026-04-23 | AIOTUSDT | 56% | +187.52% |