Forecast methodology
Read the Crypto Prediction Today ETH forecast methodology: target definition, 59 price-invariant features, frozen calibration, publication timing and forward validation.
What the model predicts
The current primary target is binary and deliberately narrow: whether ETH trades at least 4% below the next-hour forecast anchor at any point during the following 24 hours. The public number is the model's calibrated probability for that event.
59 price-invariant features
The feature vector uses normalized market information such as ETH returns, realized volatility, candle range and wick structure, ATR percentage, distance from rolling highs and lows, drawdown, breakout distance, volume z-scores, volatility-of-volatility and selected ETH/BTC relative features. Raw absolute ETH or BTC price levels are excluded from the model feature vector.
Frozen probability model
The current forward model was selected after historical research and then frozen. During the forward-validation period its weights, feature definition, target threshold and probability calibration are not updated using new outcomes.
Strict publication timing
The collector waits for the exact completed hourly signal candle. A valid forecast must be stored within the configured anchor-lag window. If the required market data arrives too late, the system records a skip rather than creating a forecast retroactively.
How outcomes are resolved
After 24 hourly candles are available, the system checks the minimum ETH low during the forecast window relative to the next-hour opening anchor. If the maximum downside excursion reaches 4% or more, the event is marked HIT; otherwise it is marked MISS. The original probability remains unchanged.
Historical versus forward evidence
Historical v1.10 out-of-sample results are shown for research context. Because earlier experimentation influenced the final target, they are not treated as independent confirmation. The forward archive is the independent evidence being accumulated now.