JumpDiffPortfolio

Portfolio allocation under a Continuous-Time Markov Decision Process (CTMDP) regime model with Jump-Diffusion dynamics, a Late-Fusion neural filter, and a Residual LSTM corrector. Built with a strict no-look-ahead discipline.

Research tool, not investment advice. Live execution (Alpaca) is gated behind explicit confirmation and defaults to paper trading.

Problem

Static mean-variance allocation ignores that markets switch between regimes (calm / stressed) with fat-tailed, jumpy returns. This project infers the latent regime online and allocates per regime, then blends — explicitly modelling jumps (Poisson-Normal mixture) instead of assuming Gaussian returns, while guaranteeing no information leaks from the future into any training or scaling step.

Data

• Prices & macro: daily OHLCV for ~17 liquid US ETFs/stocks (SPY, QQQ, sector ETFs, Treasuries, commodities, single names) from yfinance, 2010–2026, cached to Parquet under outputs/cache/. Macro series (VIX, ^TNX, DXY) are shifted one trading day at the loader to avoid look-ahead.
• Fundamentals: P/E, P/B, EPS and quarterly fundamentals (data/fundamentals.py), applied with a reporting lag so only already-published data is used.
• Evaluation is walk-forward out-of-sample: ~1,856 OOS daily observations (~7.4 years) across folds.

Approach / Model

Per-asset regime-conditioned Jump-Diffusion log-returns:

dX_i^k(t) = mu_i^k dt + sigma_i^k dW_t + J_i^k dN_t(lambda_k)

The regime k follows a CTMDP whose discrete-time posterior p_k(t) = P(z_t = k | F_t) is approximated by the Late-Fusion neural filter trained against the Jump-Diffusion mixture log-likelihood. A Residual LSTM adds a small logit-space correction p_corrected = softmax(logits_filter + delta_LSTM). The portfolio is per-regime CRRA optimisation with James-Stein shrinkage and L2 entropy regularisation, blended pi_t = sum_k p_k(t) pi_k*, projected to the simplex and risk-scaled to a target volatility.

configs/  data/(loader,splits,fundamentals)  features/(windows,standardize,macro)
models/(filter,jump_diffusion,ctmdp,trainer)  optimization/portfolio
backtest/(walk_forward,metrics,statistical_tests)  evaluation/  execution/(alpaca)
notification/  scripts/  tests/

Results (vs baseline)

Walk-forward OOS, 1,856 days aggregated across folds. Strategy metrics from outputs/metrics/aggregate_summary.csv; baselines reconstructed on the same OOS dates via python -m scripts.baseline.

Series	Sharpe	Vol	Return	Max DD
JumpDiffPortfolio (net)	0.76	15.4 %	17.0 %	−16.5 %
SPY (buy & hold)	0.40	17.4 %	11.7 %	−35.7 %
Equal-weight	0.37	13.9 %	9.6 %	−29.3 %

On point estimates the strategy beats both passive baselines on all three axes — higher Sharpe, higher return, and roughly half the drawdown (−16.5 % vs −36 %/−29 %). Sortino 1.13, Calmar 1.03, hit rate 55.4 %.

Honest significance check. Diebold-Mariano (MSE) vs the benchmark is significant (p = 0.045), but the White Reality Check p = 0.20 is NOT significant once data-snooping is accounted for, and the Sharpe 95 % CI is wide ([0.32, 1.71]). Read this as promising, not proven: the out-performance is real in-sample-of-folds but the edge is not established at the strict multiple-testing bar.

Stack

Python 3.12+ · NumPy · SciPy · pandas · PyTorch (filter + LSTM) · yfinance · Alpaca (execution) · pytest. CRRA optimisation via SLSQP (scipy.optimize).

Quickstart

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env          # rellena credenciales solo si vas a ejecutar/notificar

# Full training + walk-forward backtest (genera todas las figuras del paper)
python -m scripts.train_backtest --folds 6 --epochs 100 --lstm-epochs 80

python -m scripts.infer_daily     # inferencia diaria (sin ejecución)
python -m scripts.diagnose        # diagnósticos de sesgo (correlación vs factores)
pytest -q                         # 34 tests

Variables de entorno (solo para ejecución/notificación)

APCA_API_KEY_ID, APCA_API_SECRET_KEY, APCA_BASE_URL (paper por defecto), CONFIRM_LIVE_TRADING (debe valer yes para trading en vivo), TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID. Se leen solo del entorno; nunca hay claves en el código.

Anti-leakage discipline

1. Train/Val/Test split before any standardization or scaling.
2. Macro data shifted by one trading day (shift(1)) at the loader.
3. Sliding windows are causal (use only t-w, …, t-1).
4. The rolling standardizer fits only on the train fold.
5. Per-fold artifacts are saved only after walking forward.
6. tests/test_no_leakage.py + tests/test_core_math.py enforce these invariants.