Experiment Spec: Phase 1 — JEPA Representation PoC

Hypothesis

We believe that a TS-JEPA encoder trained on G10 FX hourly data (2008–2022) will produce latent market-state embeddings that are structurally separable by volatility regime without explicit regime labels, measurable by silhouette score > 0.35 on k-means clusters evaluated against realised-volatility regime labels on held-out 2023 data including at least one structural break.

This hypothesis is FALSE if silhouette score ≤ 0.35 OR linear probe R² ≤ 0.40 on held-out evaluation.

Prerequisites: Phase 0 passed (MAE silhouette > 0.20, TS-JEPA reproduced).

Background

Phase 0 confirmed that SSL-based representation learning can find regime structure in FX time-series. Phase 1 tests whether JEPA's specific inductive bias (predict target embeddings from context embeddings, never reconstruct raw data) produces richer representations than a simple MAE — and whether those representations are useful for risk management tasks (measurable via linear probe).

Design

Data

Source: DUKASCopy, all G10 pairs (EUR/USD, GBP/USD, USD/JPY, USD/CHF, AUD/USD, USD/CAD, NZD/USD, EUR/GBP, EUR/JPY, EUR/CHF), hourly
Train: 2008-01-01 – 2022-12-31 (all 10 pairs, jointly)
Held-out test: 2023-01-01 – 2023-12-31 (sealed until final evaluation)
Features: log-return, rolling 20-period HV, VIX (daily → hourly)
Regime label (evaluation only): rolling 30-day HV percentile; binary + 4-class (quartiles)

Model

Architecture: TS-JEPA (Ennadir et al., 2025)
- Context encoder: maps observed window → latent embedding
- Target encoder: EMA of context encoder (momentum β ≈ 0.996); stop-gradient
- Predictor: shallow MLP bridging context → target embedding
Masking horizons: ablate K ∈ {1h, 8h, 24h}; context window = 120h
Training: multi-pair joint training (one model, all G10 pairs)

Baseline

Phase 0 MAE (best masking horizon from Phase 0)
PCA on raw features (Phase 0 baseline)

Ablations

TS-JEPA vs MAE — is JEPA's no-reconstruction inductive bias adding value?
Single-pair (EUR/USD only) vs. multi-pair — does joint training improve representations?
Masking horizon K: {1h, 8h, 24h}

Acceptance Criteria

Silhouette score > 0.35 on held-out 2023 data (k-means k=3–5, binary HV label)
Linear probe R² > 0.40 on frozen embeddings vs. realised-vol decile
PC1 / rolling-HV correlation < 0.85 (encoder learning more than volatility level)
TS-JEPA silhouette exceeds Phase 0 MAE silhouette by > 5%
Rerun ×3 within ±10% of reported silhouette

Out of Scope

VaR, ES, distributional forecasting (Phase 3)
Exotic pairs beyond G10
Options pricing, alpha generation, live trading
Any data after 2022-12-31 for training; test set opened only for final evaluation
Regime detection backtesting (Phase 2 — embedding drift as early warning)

Null Result Protocol

If primary criteria not met:

If silhouette > 0.20 but ≤ 0.35: JEPA shows partial structure; not sufficient for Phase 2. Investigate whether multi-pair training, longer context window, or additional features close the gap. One retry permitted with documented rationale.
If silhouette ≤ 0.20: regression from Phase 0; investigate JEPA training stability (collapse risk). Do not proceed.
If linear probe R² ≤ 0.40 despite good silhouette: embeddings are structured but not encoding risk-relevant information. Record as a finding; reconsider feature set.
Record all results in results/summaries/phase-1-[pass|null].md
Write findings to brain: brain_write wing=jepa-fx hall=failures

Risks

Risk	Canary	Mitigation
EMA encoder collapses	All embeddings converge to near-zero; loss goes to ~0 early	Verify EMA momentum schedule; check stop-gradient implementation
Encoder learns only EUR/USD vol	PC1 dominated by EUR/USD HV even in multi-pair model	Evaluate per-pair silhouette; if EUR/USD dominates, weight loss by pair
Phase 0 silhouette was data-split artefact	Phase 1 MAE baseline doesn't reproduce Phase 0 numbers	Fix random seeds; document split methodology in Phase 0
Insufficient regime diversity (2008–2022)	Embedding clusters don't separate 2023 structural break	Verify 2023 test includes high-vol episode; add 4-class label as backup

4.3 KiB Raw Blame History Unescape Escape