BATCH_SIZE=512 per step; batched embed() at eval + export time.
78k hourly windows can't fit in GPU in one shot (was fine at 877 daily).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- train.py build(): year-based split (train≤2021, OOS≥2022) replaces
misleading 70/30 mixed-period split; true OOS val_vol_r2 now ~-0.36
vs previously reported +0.18 (artefact of cross-period data leakage)
- train.py: EXPORT_EMBEDDINGS block now exports both train+OOS embeddings
with dates and HV labels for Go eval harness
- cmd/eval: LinearProbeTrainTest uses train stats for standardisation of
both sets (no leakage); standardiseCompute/applyStandardise helpers
- internal/eval: add LinearProbeTrainTest (fit-on-train, eval-on-OOS)
alongside LinearProbe (same-set); 8/8 tests still green
Phase-0 gate result: val_vol_r2=-0.36, silhouette=0.043, erank=58.9/64.
Backbone produces high-rank embeddings (SIGReg working) but does NOT
generalize across 2021→2022 regime boundary. Gate: INCONCLUSIVE/FAIL.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Phase-0 SSL feasibility gate run on daily 2019-2023 EUR/USD (path B deviation:
not hourly 2008-2022 + Go harness as specced in #5). Results:
TS-JEPA silhouette mean=0.018 (need >0.20) — FAIL
PCA baseline silhouette=0.136 — also below threshold
sensitivity: 2000 ep + D=64 worsened to 0.004 (not a training-time issue)
Root cause: 2 daily features (ret, realized_vol) carry minimal regime structure
at this resolution. The JEPA objective with SIGReg pushes embeddings toward
isotropic Gaussian — good for downstream probes (val_vol_r2>0) but may actively
resist the clustering structure the silhouette gate measures.
Null protocol: real gate requires #4 (Go harness) + #2 (hourly data, more
features) before rerunning. HEPA (#14) noted as alternative backbone.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
All 5 kept: val_vol_r2 -0.1543 → +0.0599 (+0.214 total). Backbone learning.
Agent tuning: LR, depth, SIGREG_LAM, EPOCHS. Still well below toy ceiling
(0.37) — real backbone room to grow via #3/#4/#5.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
PatchTST-style transformer encoder with JEPA predictive loss + SIGReg
regularization (Balestriero & LeCun arXiv:2511.08544; time-series placement
from ChronoJEPA). Token-level SIGReg (dual placement) to avoid time-axis
collapse (confirmed real by ChronoJEPA). Baseline val_vol_r2=-0.1543 on first
run — expected for fresh weights with new architecture. Agent will iterate.
SIGReg source: Epps-Pulley statistic, identical math to LeJEPA MINIMAL.md.
Refs: #3 (TS-JEPA reproduce), ChronoJEPA github.com/MrRobotop/ChronoJEPA
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Toy encoder near ceiling. 1 kept (val_vol_r2 0.3032→0.3442), 5 reverts.
Consistent plateau = time to swap in TS-JEPA backbone (#3/#5).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fetch_data.py politely pulls EUR/USD M1 from histdata.com (maintained package
handles the anti-hotlink token; per-year, spaced). prepare_data.py (LOCKED per
Phase-1 contract) parses M1 -> daily series with realized_vol = the val_vol_r2
target (sqrt sum of squared intraday returns). Verified on 2019-2021: 937 days,
March-2020 COVID RV spike 5.1x over 2019 median — real signal, target works.
Data gitignored (DVC/MinIO = #10).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
scripts/check_gpu.py verifies PyTorch cu130 sees the koala Blackwell GPU
(sm_120) and computes — the Phase-0 prerequisite before any autoresearch
experiment. Verified green: torch 2.12.1+cu130, RTX 5070, GPU matmul OK.
Note: koala GPU is shared with the llama-swap LLM stack — run the autoresearch
agent on iguana/berget so the card stays free for train.py.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
mathiasandClaude Sonnet 4.6