Roadmap Overview
The plan lives at plans/260423-1037-event-conditioned-kronos/ in the repo. This page is the running overview; individual phase pages are auto-synced from the source markdown.
Three parallel tracks
Section titled “Three parallel tracks”| Track | Phases | Goal |
|---|---|---|
| A · Ensemble | Phase 0 | Add Chronos-2 as a parallel predictor for regime-shift robustness |
| B · Event-conditioned Kronos | Phases 1–5 | Extend Kronos with 20-channel event / surprise / cross-asset conditioning via LoRA |
| C · Continuity | Phase 6 | Monthly rolling fine-tune per asset class to combat distribution drift |
Tracks A and B can run in parallel. Track C is gated on the Phase 3 LoRA adapter.
Phase list
Section titled “Phase list”| # | Phase | Owner | Status |
|---|---|---|---|
| 0 | Chronos-2 Ensemble | Infra | Pending |
| 1 | Event Encoder (20 ch) | ML | Pending |
| 2 | Model Modification | ML | Pending |
| 3 | LoRA Training | ML | Pending |
| 4 | API & Batch Integration | Infra | Pending |
| 5 | Ensemble Evaluation | ML + Data | Pending |
| 6 | Rolling Fine-Tune | ML | Pending |
Decisions already locked
Section titled “Decisions already locked”- No retrain from scratch — 3+ months, no upside vs LoRA on 80 K candles.
- No Moirai-2 — CC-BY-NC licence blocker.
- Keep Kronos-base (102.3 M) on RTX 4060 — already deployed and working.
- No INT8 quantisation now — 50 ms DB cache path is not the bottleneck.
- Extend current plan rather than swap frameworks — event conditioning confirmed novel (zero forks do it per fork audit).
Success gates
Section titled “Success gates”From individual phase files, rolled up:
- Event-day spread ratio ≥ 1.5 vs baseline
- Post-surprise directional accuracy ≥ 60 %
- Non-event-day MAE within 2 % of base Kronos
- Ensemble Sharpe ≥ best single-model Sharpe + 0.1
- LoRA training completes in < 4 h on RTX 4060
- Monthly rolling fine-tune completes in < 8 h per asset class