Three studies move from representation diagnostics to a pre-registered signal scan and then a walk-forward comparison of compact and foundation-model forecasters.
Research programme
Findings so far
The usable signal is distributional rather than directional: path excursions and volatility are predictable where return sign is not. A compact from-scratch encoder offers the best quality-to-cost tradeoff, while fine-tuning helps weaker frozen representations.
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.
Report 1 · Groundwork
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.
Report 2 · Signal mapping
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.
Report 3 · Comparative study
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.