Systematic · Diversifying alpha · U.S. equities
A multi-agent quantitative investment platform: independent research pods under principal-investigator orchestration, feeding one long-only U.S. equity book. 45 signals in production, 250+ hypotheses documented, and nothing reaches the book on any agent's word alone.
Thesis
Every large multi-manager converged on the same design: independent research pods hunting alpha in parallel, with a shared data, risk, and execution platform underneath. The design's limit has always been people. There are only so many world-class pods you can field.
Polyvaria removes that limit. Its research pods are agent teams directed by centralized principal-investigator orchestration: priorities and risk budgets flow down, results and attribution flow up. A pod develops its hypotheses from idea through implementation against the shared backtester and submits the survivors for validation. The PI layer moves compute toward the families earning it and retires lines of research that stall. It also keeps the shared record of what has already failed, and a dead hypothesis doesn't get run twice. Adding a pod is a deployment rather than a hire. Research breadth grows with compute, and the gates still decide what earns weight in one book.
The platform underneath is staffed the same way. The data service, the backtester, risk, optimization, and execution are separate systems, each owned by its own agent team with its own quality rules. The separation is deliberate. A research pod cannot touch the referee that scores it or the data room it reads from, so there is no way to cheat, only to pass. Every result is recomputed from pinned inputs, and the verdict is cryptographically signed with keys no agent holds. That is what makes the breadth safe to scale. The gates don't care whether the work came from an agent or a human.
Research
It starts when the platform notices a weakness in the book: a decaying signal, a corner of the market with no coverage. An agentic principal investigator turns the gap into funded research, pods build candidate signals, and a statistical validation gate judges what they produce. Survivors are deployed and watched, the watching surfaces the next weakness, and humans govern the whole loop without sitting inside it.
Approach
The same code runs the live daily rebalance and the multi-year historical backtest. The backtest pins the as-of date and simulates its fills; everything else is the same path. Each stage is owned by one of the platform's independent agent teams.
01
Point-in-time by construction.
The investable universe is rebuilt for every historical date, delisted names included, so the backtest never sees a company the live book couldn't have traded. About 2,500 tradable names on a typical day, sixteen years deep.
point-in-time universe
02
45 signals in production.
Research pods work price, fundamental, event, analyst, and news data into cross-sectional rankings of the universe. Fundamentals enter as first filed and news as it arrived, never as later revised. The newest layer is interaction alpha: momentum confirmed by an earnings surprise, surprise confirmed by cash profitability, reversal that skips announcement windows. In testing, these conditioned signals held up where naive combinations failed. Candidates clear durability gates across three market regimes before they earn a weight.
cross-sectional alpha ranks
03
Two horizon models, one unified score.
Gradient-boosted models blend the signal panel into a unified alpha score at two- and three-week horizons. Every model is trained under a pre-registered protocol: features, folds, and acceptance gates are frozen before training starts, and every trial lands in a permanent ledger. The overfitting math never gets to forget an experiment. Refits are walk-forward with purge and embargo against leakage. In live serving, the overlay abstains rather than guesses when its inputs drift.
unified alpha score
04
Barra-grade, estimated daily.
A cross-sectional factor model in the USE4 tradition: market, seven styles (size, value, momentum, leverage, residual volatility, beta, quality), and industries, re-estimated every trading day. The covariance corrects for the ways risk models actually fail: serial correlation in factor returns, the optimizer's habit of betting hardest on the least-well-estimated directions, and volatility regimes that move faster than a trailing window. Signals are winsorized, z-scored within sector, and residualized against this structure. What survives is return the factor model can't explain, and that residual is what the book bets on.
residual alpha · factor covariance
05
Constrained conic QP.
A conic solver maximizes expected return net of risk and transaction cost, subject to long-only, sector, market-cap, single-name, and turnover constraints. Cost and turnover are charged against current holdings, which keeps the trade list stable from one rebalance to the next, and a turnover circuit-breaker sits behind the solve as a second, independent line. Each plan carries a deterministic, content-addressed identifier. Any day's solve replays exactly.
target weights · rebalance plan
06
Every order knows its urgency.
Each trade in the plan carries an urgency score built from alpha decay, risk reduction, and drift, then damped by liquidity. The score maps to a schedule: patient limit orders where the edge is slow, immediate marketable orders where it is dying. Target price bands and participation caps come from a spread-and-impact cost model, and post-trade TCA feeds realized costs back into the optimizer, and the cost curve the book trades against ends up reflecting what it actually pays.
fills · realized cost
Research Library
Each family is owned by a research pod that works it for new signals against the platform's backtester and sixteen years of point-in-time history. Published anomalies are assumed to be decaying from the day they were published, and get haircut accordingly. More candidates means more accidental discoveries, which is exactly why validation sits outside the pod. A candidate stays on the bench until the referee says it clears, and the referee is never its author.
Data
When research asks for data that doesn't exist, the platform builds it mid-investigation, and every new dataset proves it can't see the future before it lands. Most asks are answered from the catalog in seconds. What follows is the morning when the answer is no:
Verification
The platform assumes every result might be wrong until machinery says otherwise. A standing red team attacks each candidate before trial, and the verdict comes from checks its authors cannot touch.
Agent-written signal code runs in a fail-closed sandbox: no network, no filesystem, no credentials. Only numbers cross the boundary. The platform recomputes every statistic itself and signs the verdict with a key the agents don't hold, so no agent can mint its own acceptance.
Each investigation ends in a findings memo whose verdict is set by the scorer, never the author, and a row in a research log where rejects are first-class. When deeper history exposed a shipped signal as a small-sample artifact, it was deleted from production the same week. The no's are written down so they stay dead.
The linear and square-root impact coefficients in the cost model are fitted to the platform's own broker fills and refitted as trade history accumulates. Slippage assumptions track what the book actually pays to trade, and the same fills drive the post-trade TCA that order scheduling is tuned against.
Pipelines, monitors, data ingest, the backtesting harness, and the platform code itself are built and improved by the same agent pods that operate the system day to day; validation stays with gates the builders don't own. There is no separate engineering organization for research to wait on.
Controls
The agents operate the platform; people set its boundaries. The mandate, the constraint stack, and the gate thresholds belong to the platform's human principals, and nothing reaches the live book without clearing them.
Stack
The interesting risks belong in the research. The parts that touch money are chosen for determinism and auditability first, speed second, and the heavy intelligence stays upstream in the research layer.
Polyvaria is in active development and trades proprietary capital only. It is not open to outside investment or inquiries.