mathias/jepa-fx-risk
1
0
Fork 0
generated from mathias/template-go-web
Code Issues 10 Pull Requests Actions Packages Projects Releases Wiki Activity
16 Commits 3 Branches 18 Tags
fa6d6c634a2420cf87ae989f7e77c858fd16d0e3
Commit Graph
9 Commits
Author SHA1 Message Date
mathiasandClaude Sonnet 4.6 fa6d6c634a fix(train): mini-batch training to avoid GPU OOM on hourly dataset
CD / Build & Import (push) Failing after 7s
Details
CD / Deploy via GitOps (push) Has been skipped
Details
CD / Lint / Test / Vet (push) Successful in 4s
Details 
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>
 2026-06-25 13:14:29 +02:00
mathiasandClaude Sonnet 4.6 e31905dc43 feat(data): EUR/USD hourly pipeline + 2008-2023 M1 dataset (#2)
CD / Lint / Test / Vet (push) Successful in 4s
Details
CD / Build & Import (push) Failing after 8s
Details
CD / Deploy via GitOps (push) Has been skipped
Details 
- scripts/prepare_hourly.py: M1→hourly aggregation (realized_vol = sqrt(Σr²),
  MIN_BARS=30 threshold, no weekend rows, year-based split preserved)
- tests/test_prepare_hourly.py: 5 TDD tests, all green
- train.py: USE_HOURLY=True, WINDOW=240 (10-day), PATCH_LEN=24 (1-day patches);
  build() prefers eurusd_hourly.parquet, falls back to daily; EXPORT BLOCK updated
- Taskfile.yml: data:fetch:historical, data:prepare:hourly, data:prepare:all, data:test
- 98,591 hourly rows (2008-2023) covering GFC, Euro crisis, Brexit, COVID, Fed cycle

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-25 13:12:48 +02:00
mathiasandClaude Sonnet 4.6 bde651b0df feat(backbone): replace TS-JEPA+SIGReg with HEPA causal JEPA
CD / Lint / Test / Vet (push) Successful in 4s
Details
CD / Build & Import (push) Failing after 8s
Details
CD / Deploy via GitOps (push) Has been skipped
Details 
HEPA (Petersen et al., arXiv:2605.11130, ICML 2026 Spotlight):
- CausalEncoder: non-overlapping patches + per-patch LayerNorm +
  causal Transformer (generate_square_subsequent_mask) → all tokens (B, N, D)
- HorizonPredictor: MLP(cat(h_t, Δt)) → predicted future embedding;
  Δt sampled uniformly from [1, min(DELTA_T_MAX, N-1-c)] per epoch
- vicreg_loss: (1-α)·L1(norm(ĥ), norm(h*)) + α·(L_var + L_cov);
  joint training — no stop-gradient on target encoder
- Probe: last-token embedding [:, -1, :], fit on 2019-2021, eval on OOS

Results (true OOS 2022-2023):
  val_vol_r2: -0.45 (TS-JEPA+SIGReg) → +0.243/+0.276 (HEPA)
  effective_rank: 58.9/64 → 122.3/128 (near-full-rank, no collapse)
  Phase-0 gate on val_vol_r2: PASS ✓

Tests: 6/6 green (causal masking verified with non-uniform perturbation;
per-patch LayerNorm is mean-invariant so constant shifts are absorbed)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-25 08:05:33 +02:00
mathiasandClaude Sonnet 4.6 20aeecb971 fix(eval): correct probe metric to use true year-based OOS split
CD / Lint / Test / Vet (push) Successful in 4s
Details
CD / Build & Import (push) Failing after 7s
Details
CD / Deploy via GitOps (push) Has been skipped
Details 
- 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>
 2026-06-24 22:57:14 +02:00
mathiasandClaude Sonnet 4.6 7d04423d39 feat(loop): 5 iters on TS-JEPA+SIGReg backbone — consistent improvement
CD / Lint / Test / Vet (push) Failing after 2s
Details
CD / Build & Import (push) Has been skipped
Details
CD / Deploy via GitOps (push) Has been skipped
Details 
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>
 2026-06-24 07:45:55 +02:00
mathiasandClaude Sonnet 4.6 44e8b3eb95 feat(model): TS-JEPA+SIGReg backbone replaces toy encoder (#3 step 1)
CD / Lint / Test / Vet (push) Failing after 3s
Details
CD / Build & Import (push) Has been skipped
Details
CD / Deploy via GitOps (push) Has been skipped
Details 
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>
 2026-06-24 07:42:56 +02:00
mathiasandClaude Sonnet 4.6 f5ce8d6706 chore(loop): 6 more iters — plateau at ~0.34-0.37 (1/6 kept)
CD / Lint / Test / Vet (push) Failing after 4s
Details
CD / Build & Import (push) Has been skipped
Details
CD / Deploy via GitOps (push) Has been skipped
Details 
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>
 2026-06-24 07:37:29 +02:00
mathiasandClaude Sonnet 4.6 ed85dc4a8c feat(loop): 3 iterations complete — autoresearch chain end-to-end (closes #11 Phase A)
CD / Lint / Test / Vet (push) Failing after 2s
Details
CD / Build & Import (push) Has been skipped
Details
CD / Deploy via GitOps (push) Has been skipped
Details 
3 iterations ran unattended on koala (agent=berget/gemma4-31b, GPU=RTX5070):
  iter1: val_vol_r2 0.2821→0.3749 (+9.3%) KEEP  [EMBED_DIM 16→64]
  iter2: 0.2234→0.3011 (+7.8%) KEEP  [MASK_FRAC tuning]
  iter3: 0.3011→0.3032 (+0.2%) KEEP  [minor capacity tweak]
Final EMBED_DIM=64, MASK_FRAC=0.4. STATUS.md tracks full trajectory.
Both monitoring axes live: research=STATUS.md metric table, technical=GPU
snapshot per iter (0% util between runs, shared cleanly w/ llama-swap).
Phase-A acceptance: loop runs 3+ iters unattended, metric moves, both axes visible.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-24 07:17:50 +02:00
mathiasandClaude Sonnet 4.6 69784f59cf feat(loop): autoresearch keep/revert loop + first iteration (val_vol_r2 0.2821→0.3749, +9.3%) 
loop.py: Karpathy-style keep/revert loop. Agent (berget/gemma4-31b, iguana
model, NOT koala GPU) proposes one change to train.py per iter → train.py runs
on koala GPU (<2s) → read val_vol_r2 from metrics.json → keep if improved, else
restore original content. STATUS.md tracks per-iter metric + delta + GPU snap.
Iter 1 kept: improved EMBED_DIM/capacity, +9.3% on OOS R².

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-24 07:16:50 +02:00