Coordinate autonomous research agents that generate falsifiable hypotheses, validate them against history, and produce auditable research — without losing statistical discipline.
Mean reversion · vol_q1
HY OAS → R2K drawdowns
PEAD decay · post-2018
Hypotheses ship with explicit reject criteria — not vibes.
Walk-forward folds, leakage scans, and FDR control are mandatory.
Every decision an agent makes is recorded and reproducible.
Coordinate dozens of research agents across hypotheses, datasets, and experiments — with the same rigor your senior researchers apply by hand.
Scanning 412 features for IC > 0.04 in vol_q1
Hidden Markov fit · 4 states · BIC stable
Cross-asset CC matrix, lag ∈ [1,21]
Outlier days flagged: 17 · awaiting human
Mean reversion in low-vol regimes (SPX intraday)
Lead/lag: HY OAS → Russell 2000 drawdowns
Earnings PEAD decay accelerates post-2018
rejected · Coverage bias in dataset
Carry-momentum cross in 10Y rates
Across 2003–2018, daily SPX returns exhibit a statistically robust short-horizon mean reversion when realized volatility sits in the lowest quartile. Out-of-sample (2019–2024) the effect persists, with Sharpe of 1.42 net of estimated costs and a max drawdown of −7.3%.
Walk-forward folds remained directionally consistent; leakage scan flagged one candidate feature rolling_std(target) now removed. Recommended for paper promotion pending desk review.
Most agent stacks optimize for plausibility. Quantitative research demands the opposite — falsifiability. Alpha Swarm is built around that constraint.
Agents don't 'find alpha'. They draft falsifiable hypotheses with explicit assumptions, prior probability estimates, and a planned test.
Every claim is run against historical data through a validation layer that enforces train/test separation, leakage scans, and walk-forward folds.
Anti-overfitting rules — multiple-comparisons correction, regime stability, version-locked features — are part of the runtime, not a checklist.
Every decision an agent makes is recorded. Reviewers see which features, splits, and seeds produced a result — never a black box.
A single, opinionated pipeline. Each stage produces an artifact the next stage can verify — so research is composable and auditable end-to-end.
Market, fundamentals, alt-data and internal datasets — connected via MCP and immutable dataset versions.
Specialized agents propose features, regimes, lead/lag relationships, and anomalies under shared tooling.
Free-form ideas are compiled into falsifiable, parameterized hypotheses with explicit reject criteria.
Train/test separation, leakage detection, and multiple-comparisons control are enforced before a single backtest runs.
Walk-forward simulation with realistic costs, capacity assumptions, and regime-stratified performance.
Auto-drafted, fully traced memo: assumptions, splits, leakage notes, OOS results, and reviewer sign-off.
The hardest part of agentic research isn't generation — it's not fooling yourself. Alpha Swarm encodes the safeguards quant teams already trust, and applies them every time.
Sequential train/test folds across time. No information from the future leaks into the past — by construction.
Every claim ships with a held-out OOS window. The system refuses to publish a memo without it.
Every backtest, seed, and feature set is logged. Duplicate runs are detected and joined to prevent silent re-fitting.
Datasets, code, and feature definitions are content-addressed. Memos cite exact hashes — reruns are reproducible to the byte.
Benjamini–Hochberg FDR is applied across the swarm's proposals, not per agent. Volume of search is treated as a cost.
Static and runtime checks flag look-ahead features, target encoding, and survivorship — before a backtest is allowed to run.
Promotion to paper or live trading requires a human reviewer to sign the trace. The runtime won't bypass it.
Performance is reported per regime (vol, rates, dispersion). Strategies that work only in one regime are surfaced, not buried.
Alpha Swarm is currently onboarding a small number of design partners. If you run a research team and care about discipline as much as discovery, we’d like to talk.