The leakage premium in decomposition-based cryptocurrency forecasting

Abstract Decomposition-ensemble models, which split a price series into components with methods such as CEEMDAN and VMD and forecast each component with a neural network, report remarkable accuracy for cryptocurrency returns. This paper shows that most of that accuracy is look-ahead. The decompositions are two-sided, so components computed on the full sample carry future prices into every input. A typical pipeline that stacks VMD on top of CEEMDAN is run on daily prices of five major crypto-assets from 2018 to 2026 under three protocols that differ only in the information each decomposition uses, with data, networks and random seeds held fixed. With the full-sample decomposition the model calls the direction of next-day returns correctly on 93 per cent of days and explains 96 per cent of their variance. Computed causally, the same model is right on 52 per cent of days and has an out-of-sample R 2 of -23 per cent, worse than a forecast of zero. Look-ahead accounts for essentially the entire premium in directional accuracy, a premium that grows with each added decomposition layer and hardly moves across market regimes. A second mechanism explains why causal decomposition forecasts are harmful rather than merely uninformative. The revisions of end-point components cancel across components in the target but not in the forecasts. A long-flat rule driven by the leaky forecasts would show phantom profits of about 146 basis points a day net of costs. A simple falsification test, proposed here, detects look-ahead in any such pipeline. JEL classification: C45; C53; C58; G14; G17