Guides

Reproduction is the only test that separates a result from a story, and most quant papers fail it somewhere between data alignment and cost assumptions. A six-stage working checklist — acquire, align, reimplement, reconcile, stress, document — with the traps at each stage and a downloadable version.

Generic ML tracking optimizes for finding the best run; quant tracking must also count every run — because the trial count is a statistical input, not bookkeeping. What to record, the finance-specific fields generic tools miss, and a minimal stack: one tracked-run decorator over one database.

The 821 papers in this area of the archive, organized into 5 branches with the strongest work in each ranked by rigor-weighted score. Regenerated as the pipeline scores new papers.

The 644 papers in this area of the archive, organized into 5 branches with the strongest work in each ranked by rigor-weighted score. Regenerated as the pipeline scores new papers.

The 1776 papers in this area of the archive, organized into 5 branches with the strongest work in each ranked by rigor-weighted score. Regenerated as the pipeline scores new papers.

The 630 papers in this area of the archive, organized into 5 branches with the strongest work in each ranked by rigor-weighted score. Regenerated as the pipeline scores new papers.

The 331 papers in this area of the archive, organized into 5 branches with the strongest work in each ranked by rigor-weighted score. Regenerated as the pipeline scores new papers.

The 93 papers in this area of the archive, organized into 5 branches with the strongest work in each ranked by rigor-weighted score. Regenerated as the pipeline scores new papers.

The 762 papers in this area of the archive, organized into 5 branches with the strongest work in each ranked by rigor-weighted score. Regenerated as the pipeline scores new papers.

Every return stream is a mix of market exposure (beta), systematic style exposures (alternative risk premia), and whatever’s left (alpha) — and the mix, not the headline return, determines what a strategy is worth and what it should cost. The sorting regression, the migration of alpha into beta, and the diligence questions.

Every strategy has a size at which its edge equals its impact — and almost no paper reports it. The estimation arithmetic from ADV and turnover, the capacity hierarchy from HFT to low-turnover factors, and the production reality that capacity is consumed by everyone running the trade, not just you.

CSCV asks a brutally direct question: across many symmetric in-sample/out-of-sample splits, how often does your in-sample winner underperform out of sample? The construction, working pseudocode, interpretation thresholds for PBO, and what the method cannot catch.

Every backtest sits on an adjustment policy, usually somebody else’s and usually undocumented. How split and dividend adjustments actually work, why price-only adjustment understates returns by the dividend yield, the actions that break naive pipelines, and the architecture that makes it all auditable.

Relative-strength momentum (rank assets, long winners, short losers) and trend-following momentum (each asset vs its own history) share a name and a lookback window — and differ in beta, crash mode, evidence base, and capacity. The comparison, mechanics-first, with each family’s production failure catalog.

‘Momentum works’ is two unrelated claims wearing one name: relative ranking predicts relative returns (cross-sectional), or a series’ own past predicts its future (time-series). They differ in hedging, sample size, crash modes, and evidence standards — and conflating them produces phantom strategies.

Event-driven research lives or dies on three questions papers rush past: was the event knowable as defined, when exactly did the information arrive, and could anyone trade the window being measured? The event-study mechanics, the leakage taxonomy specific to events, and the execution reality.

Free market data is not a worse version of paid data — it’s a different object missing specific properties: delisted history, point-in-time integrity, licensing rights, and support guarantees. The six dimensions that actually separate the tiers, where free genuinely suffices, and the upgrade triggers.

Individual studies find results; pipelines compound them. The seven stages of a repeatable research process — each with the artifact it produces and the gate it must pass — plus the experiment log that makes your trial count honest and the infrastructure map for running it all.

Replication time is your scarcest research resource; most papers don’t deserve it and a few repay it many times over. The expected-information framework, five selection criteria with a scoring rubric, and the replication-portfolio approach that beats picking one ‘best’ paper.

Most good papers are not tradeable and most tradeable insights are not publishable — the incentives point different directions. Five tests separate academic contribution from implementable idea, and knowing which you’re holding changes what you do with it.