docs: add DECISIONS.md with architecture and data decisions

2026-05-28 05:41:59 +00:00
parent d1193a57c6
commit a0d180e37d
1 changed files with 215 additions and 0 deletions
@@ -0,0 +1,215 @@
 # Architecture Decision Records
 This file records significant technical and research decisions for `jepa-fx-risk`.
 Each record is immutable once merged — append new records rather than editing old ones.
 Format: ID · Date · Status · Context · Decision · Rationale · Consequences.
 ---
 ## ADR-001 · Architecture: TS-JEPA + SIGReg as Phase 1 backbone
 **Date:** 2026-05-28
 **Status:** Accepted
 **Supersedes:** informal decision to use TS-JEPA standalone (pre-ADR)
 ### Context
 Four JEPA variants were evaluated for FX volatility forecasting and VaR/CVaR estimation:
 | Variant | Origin | Key property |
 |---|---|---|
 | TS-JEPA | Ennadir et al., Sep 2025 | Time-series native; EMA collapse prevention |
 | LeJEPA | Balestriero & LeCun, Nov 2025 | Proven optimal embeddings (isotropic Gaussian); SIGReg |
 | MTS-JEPA | He et al., Feb 2026 | Multi-resolution + codebook; no public code |
 | Var-JEPA | Multiple, Mar 2026 | ELBO-based UQ; no public code |
 Key constraints: 12 GB VRAM (Blackwell, koala), hourly DUKASCopy data, internal PoC target,
 autoresearch loop requires a clean single-scalar search space, trading desk requires an
 explainable theoretical story.
 ### Decision
 Use **TS-JEPA architecture with SIGReg replacing EMA** as the Phase 1 backbone.
 Concretely:
 - Start from the TS-JEPA open-source implementation (arXiv:2509.25449, GitHub)
 - Remove the EMA target-network mechanism
 - Replace it with Sketched Isotropic Gaussian Regularization (SIGReg) from LeJEPA
  (arXiv:2511.08544), controlled by a single λ hyperparameter
 - Keep TS-JEPA's temporal patchwise masking and Transformer encoder unchanged
 ### Rationale
 **Why not pure TS-JEPA:** EMA is a heuristic; λ interacts with EMA momentum and
 learning rate, creating a three-way search space that is hard to navigate with autoresearch.
 EMA also has no theoretical non-stationarity guarantee.
 **Why not pure LeJEPA:** The reference implementation targets vision (multi-crop views).
 Adapting it to temporal patchwise masking requires non-trivial surgery and moves away from
 open code. TS-JEPA's masking is already the right inductive bias for time series.
 **Why the hybrid:** SIGReg is architecture-agnostic — it operates on the embedding
 distribution, not the encoder structure. Swapping EMA for SIGReg is a ~20-line change to
 TS-JEPA's training loop. The result is:
 - Time-series native (TS-JEPA masking + patch structure)
 - Provably collapse-free without heuristics (SIGReg)
 - Single search axis for autoresearch (λ ∈ [0.01, 1.0])
 - Non-stationarity robustness proven formally (arXiv:2602.19373 extends LeJEPA
  guarantees to non-stationary target distributions — directly relevant to FX)
 - Explainable to a model validation team: "embeddings are provably optimal for
  downstream prediction under distributional uncertainty"
 **Why not MTS-JEPA or Var-JEPA now:** Both lack public code (as of May 2026).
 MTS-JEPA's multi-resolution objective is the right next hypothesis (see ADR-003).
 Var-JEPA's ELBO-based UQ is a compelling future direction for CVaR estimation.
 ### Consequences
 - Phase 0 (MAE baseline) is unaffected — it precedes the JEPA architecture choice
 - Issue #3 (TS-JEPA reproduction) is still the right first step; SIGReg is added after
  reproduction is confirmed
 - The autoresearch `program.md` primary search axis is λ (SIGReg weight)
 - Secondary axes: masking block size, patch stride, encoder depth
 - `model/requirements.txt` must include the SIGReg implementation (≈20 lines,
  can be vendored directly)
 ---
 ## ADR-002 · Data: DUKASCopy hourly G10 FX as primary training data
 **Date:** 2026-05-28
 **Status:** Accepted
 ### Context
 Data scale is the most dangerous assumption for any SSL/JEPA approach. Daily FX data
 (~5,000 samples over 20 years) is insufficient for self-supervised pretraining.
 Two alternatives were considered: daily public data (yfinance) vs. hourly tick data
 (DUKASCopy, free, rate-limited HTTP API).
 ### Decision
 Use **DUKASCopy hourly OHLCV** as the primary data source.
 - 10 G10 pairs: EURUSD, GBPUSD, USDJPY, USDCHF, AUDUSD, NZDUSD, USDCAD,
  EURGBP, EURJPY, GBPJPY
 - Training window: 2008-01-01 – 2022-12-31 (~175,000 samples per pair)
 - Validation window: 2023-01-01 – 2023-12-31 (~2,600 samples)
 - Test window: 2024-01-01 – 2024-12-31 (held out, never seen during development)
 - Features per bar: log-return, log rolling-20-period HV, VIX (daily interpolated)
 - Weekend gaps handled explicitly — no interpolation across market close
 ### Rationale
 Hourly data gives ~35× more samples than daily. This is the minimum threshold for
 JEPA-style SSL to show a training signal within 10-minute autoresearch experiments.
 DUKASCopy is free, reliable, and provides consistent tick-level source data back to 2003.
 ### Consequences
 - The Go data pipeline (Issue #2) is the critical path for everything else
 - Phase 0 MAE baseline trains on the same 2008-2022 window
 - Daily data (yfinance) may still be used for VIX and rate differentials as auxiliary features
 ---
 ## ADR-003 · Research roadmap: Phase structure and JEPA variant progression
 **Date:** 2026-05-28
 **Status:** Accepted
 ### Decision
 Three-phase research roadmap:
 **Phase 0 — SSL feasibility gate (MAE baseline)**
 Implement a 1D temporal MAE (not JEPA) on EUR/USD hourly 2008-2022.
 Gate criteria: silhouette > 0.20 on 2023 held-out, MAE > PCA baseline, ±10% over 3 reruns.
 Purpose: validate that the data and eval harness work before committing to JEPA complexity.
 If gate fails: follow null result protocol in `specs/phase-0-ssl-feasibility.md`.
 **Phase 1 — TS-JEPA + SIGReg autoresearch sweep**
 Primary architecture per ADR-001.
 Autoresearch loop: `program.md`-driven, 10-min experiments, 50-experiment budget.
 Primary metric: `val_vol_r2` (linear probe R² on 1-day realized volatility).
 Gate criteria: `val_vol_r2` > GARCH-implied baseline AND Kupiec p-value > 0.05 on
 EUR/USD VaR 99%.
 Kupiec is logged from experiment 1 to verify it co-moves with `val_vol_r2`.
 **Phase 2 — MTS-JEPA multi-resolution hypothesis**
 Introduce parallel multi-scale predictive pathways (1h, 8h, 24h context windows)
 adapted from MTS-JEPA (arXiv:2602.04643).
 Hypothesis: multi-scale representations improve regime detection (silhouette) and
 reduce VaR exceedance clustering (Christoffersen test).
 Prerequisite: Phase 1 gate passed AND MTS-JEPA code available or reproducible from paper.
 Time-box: if MTS-JEPA code not available within 4 weeks of Phase 2 start, implement
 multi-resolution masking from scratch using Phase 1 backbone as base.
 **Phase 3 — Internal bank data (future)**
 Replace DUKASCopy pipeline with internal tick feed adapter.
 Fine-tune heads only; backbone frozen or lightly fine-tuned.
 Out of scope for current PoC cycle.
 ### Consequences
 - Issue #5 (Phase 0 MAE) is the unblocked next executable step
 - Phase 1 autoresearch is blocked until Phase 0 passes its gate
 - Var-JEPA (ELBO-based UQ) is a named future hypothesis for CVaR estimation in Phase 2+
  but not on the critical path
 ---
 ## ADR-004 · Evaluation: Go harness + Python training separation
 **Date:** 2026-05-28
 **Status:** Accepted
 ### Decision
 Hard separation between training (Python) and evaluation (Go):
 - **Python** (`model/`): all training, embedding export, model checkpointing
 - **Go** (`src/eval/`): all evaluation metrics — silhouette, linear probe R², collapse
  diagnostic, Kupiec/Christoffersen backtests
 - Interface: Python exports embedding matrices + labels to `experiments/RUNID/` as
  `.npy` files; Go eval harness reads them and writes `metrics.json`
 ### Rationale
 Go evaluation gives deterministic, fast, auditable metric computation with proper
 unit tests. It decouples the experimental loop from the training framework, making
 it possible to re-evaluate any past experiment without re-running training.
 The Go layer also serves as the foundation for the eventual trading desk dashboard.
 ### Consequences
 - All acceptance criteria in Issues #4 and #5 are specified in terms of Go eval outputs
 - `val_vol_r2` (the autoresearch optimization metric) is computed by the Go harness,
  not inside the Python training loop
 - Python training loop calls `task eval:probe` as a subprocess after each experiment
  to get the scalar fed back to autoresearch
 ---
 ## ADR-005 · Compute: Blackwell GPU on koala, PyTorch cu130
 **Date:** 2026-05-28
 **Status:** Accepted
 ### Decision
 All GPU training runs on koala (Arch Linux, Blackwell GPU, 12 GB VRAM).
 PyTorch install: `pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130`
 (CUDA 13.0 wheel — required for sm_120 Blackwell support; stable as of May 2026).
 Driver requirement: NVIDIA R570+, CUDA toolkit 12.8+.
 Ollama on iguana (Mac Studio M2 Ultra) serves the autoresearch agent LLM via the
 existing LiteLLM proxy on piguard. Agent calls never hit koala directly.
 ### Consequences
 - `model/requirements.txt` must NOT pin torch to a cu124 or earlier wheel
 - CI (Issue #7) must NOT run GPU tests — CPU-only for unit tests, GPU only via
  `task experiment:run` on koala
 - 12 GB VRAM is sufficient for <5M parameter models at batch=64; monitor if
  autoresearch explores larger architectures