Research series
A staged programme on forecasting BTCUSDT perpetual futures distributionally — asking where a volatile market actually carries learnable signal before committing a model to it.
Three studies in sequence: groundwork establishing that frozen foundation-model embeddings are directionally blind but volatility-informative, a pre-registered scan mapping where signal exists across timeframes and horizons, and the completed walk-forward comparison staged on those verdicts. Each entry below opens its full report.
The theme
Short-horizon crypto forecasting rewards the wrong thing: directional edges evaporate out of sample, and validation wins vanish on test. This programme forecasts BTCUSDT perpetual futures distributionally — conditional quantiles of returns, path excursions, and volatility — and decides where and what to model from pre-registered evidence rather than architectural intuition.
Direction at these horizons barely beats a coin flip, but path excursions and volatility are strongly predictable. Conditional quantiles of how far price travels for and against a position are exactly what take-profit and stop-loss placement consume.
Verdict thresholds are fixed before the data is scanned, and every claim must hold on validation and test jointly — the guard against the validation-only mirages that plague market machine learning.
The groundwork establishes what frozen foundation-model embeddings can and cannot see; the scan maps where learnable signal exists; the comparative study stages the backbone comparison on those cells.
The arc
Three studies in sequence, each feeding the next: groundwork that redirected the objective from direction to scale, a scan that located where the signal lives, and the comparison staged on what it found. Each entry opens its full report.
Head-and-loss research on cached frozen Moirai-MoE embeddings of BTCUSDT, prompted by a production head that settled on a near-constant median. Its finding shapes everything after it: the predictable structure is scale rather than direction, and permutation controls place that skill in the causal volatility regime rather than in the backbone.
A pre-registered scan of the full BTCUSDT store asking where learnable signal exists at all. Verdict: the signal is distributional — excursion quantiles and volatility, not direction — strongest at fine timeframes and decaying with horizon. Its verdicts set the parent study's targets, scales, and horizons.
The completed walk-forward comparison staged on the scan's verdicts: a two-million-parameter from-scratch multi-resolution encoder against frozen and LoRA-fine-tuned Chronos-T5 and Moirai-MoE lanes. The small scratch encoder is the quality-versus-cost frontier, and fine-tuning helps exactly in proportion to a frozen representation's deficit.