‘Edge is feature engineering’
A recent quant panel argued the real edge today is how teams structure features and integrate regime detection—not the choice of model class—calling out regime-aware feature pipelines as the differentiator. That framing shifts project value from model novelty to robustness and feature design. (youtube.com)
The Delivering Alpha panel on November 14, 2025 featured Leda Braga (Systematica) and Afsaneh Beschloss (RockCreek), was moderated by Scott Wapner, and framed feature engineering and macro-regime interpretation as central to institutional alpha-generation amid Systematica’s ~\$12B AUM and RockCreek’s ~\$17B AUM. (hedgefundalpha.com) An academic implementation called RegimeFolio (Yiyao Zhang et al., arXiv, 14 Sep 2025) used a VIX-based regime classifier, regime- and sector-specific ensemble learners, and a dynamic mean–variance optimizer across 34 large-cap U.S. equities (2020–2024), reporting a 137% cumulative return, a Sharpe ratio of 1.17, 12% lower max drawdown, and 15–20% improvement in forecast accuracy versus benchmarks. (arxiv.org) A peer-reviewed pipeline, WaveESN–RegimeMLP (Computational Economics, published 21 Mar 2026), combined wavelet multiscale features, an echo-state reservoir with elastic-net readout, and a residual-based regime detector on AAPL, DIA, SPY and JPM (1 Jan 2021–1 Jan 2022), cutting one‑day‑ahead RMSE by 10–37% and z‑normalized RMSE by 10–32% relative to a reservoir baseline with the largest gains in volatile periods. (link.springer.com) Vendor tooling and startups have productized regime detection and feature pipelines—Quant HQ launched an API-first regime-detection and stock-ranking stack in mid‑2025, and Quantreo publishes notebooks and tutorials that embed regime-aware feature engineering into production workflows. (quanthq.io) Practitioner engineering blueprints and open-source projects demonstrate the pipeline split used in production: Mahad Khanleghari’s quant system architecture explicitly separates ingestion, feature engineering, regime detection, alpha generation and portfolio construction, while GitHub repos such as Prat1331 implement regime-aware intraday prediction with classical ML and feature modules. (mahadkhanleghari.github.io) Industry primers and newsletters list the concrete methods in use—Hidden Markov Models for latent-state inference (PyQuant), HMM+RandomForest walk‑forward recipes (QuantInsti’s Python guide), and distributional drift metrics like Wasserstein distance for regime change monitoring in practitioner labs. (pyquantnews.com)
