How it works.
Honest headline: Does not beat trading fees in testing. In walk-forward testing the model beat fees. On six months of data it had never seen, it found nothing worth trading and made 0 trades. We publish its calls anyway, because a live record made before outcomes exist cannot be tuned to those outcomes after the fact. It can still be cherry-picked by how long we run it or which version we report, which is why every version and every receipt stays up.
What it predicts
Every hour, for BTC, ETH and SOL, the model estimates two probabilities: that the price ends the next 12 hours more than 1% higher (P(big up)) and more than 1% lower (P(big down)). A move that small would be eaten by fees, so smaller moves count as nothing happening. The call is UP when P(big up) minus P(big down) is above 0.16, DOWN when it is below minus 0.16, and NO CALL otherwise. The Hopium Meter shows that difference as a lean from 50% to 100%; it is not a probability.
Entry is the close of the last finished Binance 1h candle when the call is made. Exit is the close of the candle that ends 12 hours later. A call is a hit if the price moved its way, and dead money if it moved less than 0.1% either way. The live price ticker on the site is decoration and is never used for grading.
Where the data comes from
| source | host | use |
|---|---|---|
| Binance spot 1h candles | data-api.binance.vision | Live. Every price the model sees, every entry and every exit. |
| Deribit perpetual funding | deribit.com public API | Live. What leveraged traders pay to hold positions, hourly. |
| Fear & Greed index | alternative.me | Live. Yesterday's value only, so it is always already published. |
| Coinbase candles | Coinbase Exchange | Backup for rebuilding history. Never used for live calls or grading. |
| Binance futures (funding, open interest) | fapi.binance.com | Not used. Geo-blocked (HTTP 451) from the US servers that run the model. |
| Bybit | api.bybit.com | Not used. Blocked (HTTP 403) from the same servers. |
The rule: a number can only be a model input if the hourly job can fetch it live, every hour, from where it runs. A feature that exists in training but is stale or missing in production makes the backtest a lie.
What it looks at
Recent returns (1 to 168 hours), RSI, MACD, ATR, Bollinger %B, distance from VWAP, volume and trade-count surprises, the share of volume from aggressive buyers, distance from daily and weekly highs and lows, realised volatility and its percentile, how ETH and SOL move relative to BTC, funding rates, Fear & Greed, hour, weekday and trading session. A test changes every future candle, funding value and index reading at random and checks that no past feature moves by even one bit.
How it was tested
Walk-forward only: train on the past, test on the next unseen month, roll forward, 20 months in total (2024-09..2026-04). Inside each training window the last 90 days choose the trade threshold, so no test month picks its own threshold. But the config itself was chosen by comparing walk-forward results across 39 configurations (all logged, losers included), so the walk-forward figure is a selected result, not an untouched test. The most recent six months were locked away before any of that and used exactly once, on the single chosen config.
What 91.0% means, exactly
Backtest simulation, test months 2024-09-01 to 2026-04-30 UTC (dev period); holdout 2026-05-01 to 2026-11-03 UTC.
- Compounded return of a simulated equal-weight portfolio: three fixed sleeves of 1/3 of capital (BTC, ETH, SOL), each compounding its own trades, never rebalanced, no leverage. It is not a sum of trade returns (that sum is 201.0%).
- At most one open trade per coin; signals during an open trade are skipped, so capital is never used twice.
- Entry at the reference candle close the forecast was computed from, exit at the close 12h later. That is the forecast-grading convention, not an executable fill: a real order could only be placed after the forecast existed.
- Costs: 0.2% per round trip only. Not modelled: slippage, spread, short borrow or funding, taxes.
- Shorts: return = minus the move, as on a linear perpetual with zero funding.
- Sharpe: daily equity, annualised with sqrt(365), 95% interval from a 20-day block bootstrap (1,000 draws). Not adjusted for picking the best of 39 configs.
- Max drawdown on the daily equity curve, which books each trade on its entry day, so intratrade swings are not seen.
- Execution-convention (post-publication entry) walk-forward: not computed. Needs the research candle history; marked unavailable rather than estimated.
- walk-forward trades
- 988
- compounded 3-sleeve net, 0.2% fees
- +91.0%
- Sharpe (95% CI)
- 1.27 (0.02 to 2.50)
- max drawdown
- 23.1%
- win rate
- 49.8%
- avg win / loss
- +2.25% / −1.83%
- holdout trades
- 0
- holdout net
- 0.0%
Nearly all of the gain came between December 2024 and April 2025. From January 2026 the model mostly declined to trade.
Why fees eat the edge
The first version tried to call every up or down move. It was right 52.8% of the time over 45,259 hours, which is a real edge. But the average trade earned 0.013% before fees, and buying and selling costs 0.2%. Following it lost 92.6% of the money in simulation. Being right slightly more often than a coin is easy. Being right by more than the toll booth charges is the whole game.
Why we distrust 90%
Short-term crypto prices are close to a random walk. A claim of 90% directional accuracy on hourly-to-daily crypto moves is far more often explained by grading after the fact, counting only chosen calls, or leaking tomorrow into today than by real skill. Our own rule: if a backtest scores above 65%, we treat it as a probable bug and stop. Our models have landed between 50% and 55%; that is our experience, not a law. And accuracy alone says little: a call can be right on direction and still lose money after fees.
What it cannot do
It knows nothing about news, hacks, regulation, ETF flows or anyone's tweets. It has seen about three years of one market cycle. Its edge in testing came from one stretch of 2024 and 2025 and faded. It never trades, never holds your money and never tells you how much to bet. If you act on a call, that is your decision and your risk.
Status
- model version
- v3-12h-logreg-20261104
- last retrain
- last successful prediction run
- last grading run
- missed runs, last 7 days
- 0
- errors, last 7 days
- 2
Full numbers, every configuration and the code: RESULTS.md. Or read the two-minute case study.