A Master's-thesis research programme on serial autocorrelation and trend-persistence in linear derivatives. MSc 2 Mémoire, INSEEC — Lucas Joly.
This repository is the curated, presentation-facing extract of a ~36-year systematic trading research project: a from-scratch C++/Python backtesting engine, a library of 45 individually-tested alpha signals, a flagship multi-alpha strategy validated with an institutional statistics battery, and a GAN-based synthetic-market robustness pipeline.
Research question. Can a systematic multi-alpha futures strategy — where every parameter is derived from first principles or sourced from the academic literature, with zero fitted parameters — deliver a return premium that survives the most adversarial multiple-testing correction?
Note. This is a read-only showcase. The source is shared for review, not execution: the proprietary NorgateData price panel and the data-loading layer (
shared_config,Data/) are private, so the scripts here reference them but do not ship them. Read the code and results — don't expectpython …to run end-to-end.
A four-alpha, equal-weight composite (Time-Series Trend · Carry · Skew · Vol-of-Vol) on a frozen universe of 62 liquid futures, backtested day-by-day over 1990–2026 (≈9,341 trading days) on the Carver (2023) forecast-to-position framework.
| Metric | Value |
|---|---|
| Raw annualised Sharpe (excess of T-Bill) | 0.97 |
| BHY-corrected Sharpe (Harvey–Liu, K = 118 audited trials) | 0.78 |
| CAGR | 15.5% |
| Annualised volatility | 12.9% |
| Max drawdown | 21.6% |
| Calmar | 0.72 |
| Fama–French 5-factor alpha | 10.6% (t = 3.21, p = 0.0013) |
| Fung–Hsieh 7-factor alpha | 13.0% (t = 5.83) |
| Decade Sharpes (90s / 00s / 10s / 20s) | 0.90 / 1.02 / 1.19 / 0.74 |
| Performance retained under T+1 execution lag | 98.1% |
Why it's honest. S183 has zero parameters optimised on the backtest — each value is
either a closed-form derivation (e.g. FDM_CAP = √4 = 2.0, sigmoid steepness = 2/vol_target)
or a literature citation (EWMAC speeds from Levine & Pedersen 2016; Vol-of-Vol direction
from Baltussen et al. 2018). Robustness is established across a Deflated-Sharpe test
(z = 4.28), a 1,000-draw Monte-Carlo parameter search (realised SR in the top 1.8%),
signal-shuffle null tests (24–33σ above null), and a WGAN synthetic-market stress test.
The strategy was run forward on 145 trading days of genuinely out-of-sample data (2 Jan – 26 May 2026) that did not exist when the model was frozen — a live-with-fake-money proof of concept, not a backtest.
| Metric (145 days) | Value |
|---|---|
| Sharpe (excess of T-Bill) | +0.44 |
| Annualised return | +7.33% |
| Annualised volatility | 8.56% (well below the 20% target — conservative by design) |
| Max drawdown | −4.33% |
Honest read. The sample is far too short for statistical significance. In this strongly risk-on window the strategy underperformed equity and CTA benchmarks on return, dragged by a synchronous trend reversal in agriculturals — but it held the smallest drawdown of the peer set, and its return distribution (skew, kurtosis) matched the 36-year backtest almost exactly at a lower scale. Against 10,000 bootstrapped historical 145-day windows the realised result sits around the 33rd percentile on return/Sharpe. A fuller update is scheduled for the September 2026 defence; meaningful live conclusions need 2–3 years of data.
| Folder | What's inside |
|---|---|
thesis/ |
Compiled thesis (French corpus + annexes, full English translation) and the S183 technical report (PDF). |
engine/ |
The backtesting framework: shared compounding engine, a C++ (pybind11) accelerator for the day-by-day walk-forward, a Gaussian-HMM regime detector, and a Bayesian (NumPyro/JAX) latent-factor risk filter. |
flagship_S183/ |
The flagship strategy source, its 1,000-line technical report, result figures, and the robustness-run scripts (half-life sweep, liquidity-weighting, signal-shuffle, capacity analysis). |
alpha_library/ |
45 individually-backtested alpha signals, each a standalone module, plus the cross-alpha correlation matrix and performance summary. |
gan_robustness/ |
A conditional-WGAN pipeline that trains a generative model on the real market panel and re-runs the strategy on synthetic paths to test for backtest overfitting. |
production/ |
A lean reimplementation of the S183 signal and portfolio path for daily paper trading — four dependencies, no pandas, with a differential test suite that verifies each stage against the thesis implementation. |
45 signals were each isolated, backtested over the full 1990–2026 window, and ranked. The strongest standalone sleeves (below) are the raw material for the S183 composite; the point of the exercise is that the combination — not any single signal — is what survives correction.
| Top standalone alphas | Full-sample Sharpe | CAGR |
|---|---|---|
| EWMAC Ensemble (trend) | 0.57 | 14.3% |
| EWMAC Slow | 0.57 | 14.2% |
| Normalised Momentum | 0.56 | 14.4% |
| Carry | 0.48 | 9.6% |
| Skew | 0.48 | 10.6% |
Full ranking of all 45 in alpha_library/single_alpha_summary.csv.
- Data — ~110 continuous futures (Panama-adjusted) from NorgateData across equities, rates, bonds, FX, commodities, volatility, carbon and crypto; frozen universe, strict 5-year per-instrument out-of-sample gate, no survivorship bias.
- Engine — vectorised NumPy signal generation → shared date-alignment → day-by-day walk-forward P&L and compounding (C++ accelerated) → geometric metrics. Position sizing via Carver forecast-to-position with a universe-pooled 4×4 Forecast Diversification Multiplier and sigmoid vol/drawdown overlays.
- Statistics — Newey–West HAC Sharpe SEs, Jobson–Korkie–Memmel paired tests, Ledoit–Wolf block bootstrap, Deflated Sharpe (Bailey & López de Prado), Harvey–Liu–Zhu multiple-testing haircut, Fama–French 5F / Fung–Hsieh 7F factor regressions.
- Tooling — Python (NumPy, Pandas), C++ (pybind11), PyTorch (WGAN), NumPyro/JAX (Bayesian filter), LaTeX (thesis).
- Mémoire — Corpus · Annexes (French)
- Full English translation
- S183 technical report — the complete strategy write-up, testing suite, and adversarial-critique section.
Title: “Autocorrélation sérielle et phénomènes de persistance des tendances sur les dérivés linéaires.” Tuteur de mémoire : Georges Lionel.
Lucas Joly — MSc 2 Finance, INSEEC. LinkedIn · [email protected]
Research self-financed. This repo is a curated extract of a much larger private research codebase; it is shared for academic and professional review.