Long‑only · systematic
The architecture of a top multi-manager fund, operated by self-coordinating agent teams. Forty-two uncorrelated signals running in production today.
Thesis
The world's largest quant funds and multi-manager platforms are built the same way: dozens of independent research pods hunting alpha, with a deep back office of data, infrastructure, and risk underneath. Polyvaria runs that same architecture, with self-coordinating agent teams in the seats.
Each research pod continuously evaluates new, uncorrelated sources of alpha, and adding another pod is a deployment rather than a hire. The data onboarding, backtesting, risk-modeling, and transaction-cost-analysis stack underneath the research pods is built and validated by agents too, so the research backlog never bottlenecks behind a finite engineering bench. This is not speculative: OpenAI's reasoning models recently disproved a longstanding conjecture in discrete geometry, demonstrating research-grade capability that did not exist a year ago. Aimed at a research and platform backlog, that capability outpaces the engineering throughput of even the largest funds.
Approach
The same code runs the live daily rebalance and the multi-year historical backtest. As-of date is the only thing that changes between them.
Survivorship-corrected.
The investable universe is reconstructed for every historical date from a delisted-companies snapshot. The platform never trades names it could not have known about at the time.
Forty-two uncorrelated, durable signals.
A research pipeline ingests price, fundamental, event, and news data and produces cross-sectional rankings of the universe. Signal families span value, momentum, quality, capital discipline, defensive, reversal, liquidity, post-earnings drift, and sentiment. Every candidate has to clear an orthogonality and decay-resistance bar before it earns a weight in the live book.
Conditional signal weighting.
A regime classifier across volatility state, yield-curve shape, growth/inflation quadrant, and breadth maps the current market to a weighting overlay. Momentum, defensive, value, and reversal families lean in or out of the book as the regime shifts, rather than being statically blended.
Five-factor structural, rolling exposures.
Market, size, value, profitability, and investment exposures are estimated with rolling regression. Raw signals are winsorized, z-scored within sector, and residualized against the factor structure, leaving an expected-return vector orthogonal to known risk premia. The same factor covariance enters the optimizer's objective directly, with shrinkage on residual variance.
Constrained mean-variance, warm-started.
A conic solver maximizes expected return net of risk and transaction cost, subject to long-only, sector, market-cap, single-name, and turnover constraints. The daily solve warm-starts from the previous one, keeping the trade list stable across rebalances.
Cost-aware, routed, measured.
A scheduling layer slices parent orders across the day against a spread-and-impact model, with participation caps and child-order routing tuned to venue liquidity. Post-trade TCA feeds realized cost back into the optimizer's transaction-cost term. The cost curve the platform trades against tomorrow is calibrated on what it actually paid yesterday.
Research Library
A team of research agents owns each family, mining its domain for high-IC signals against the platform's backtester and historical archive. A candidate only earns a live weight after clearing an IC bar on a holdout window.
Principles
Every layer of the stack has a feedback loop attached to it. As the underlying models improve, those loops close faster and the platform sharpens with them.
Every cached signal is evaluated daily against its information coefficient over the prediction horizon. Decaying signals are flagged before they earn weight. The machine-learning overlay refits on a walk-forward schedule with an embargo that prevents leakage, and each evaluation cycle adds new hypotheses to the cache and shifts the weights on the old ones.
Factor model, exposure estimator, covariance assembler, constraints, cost model: each component is hot-swappable from configuration. Backtest-replay tooling lets agent pods run counterfactual configurations against history and reports which parameter and component choices move attribution.
Linear and square-root impact coefficients in the cost model are calibrated against realized broker fills. Holdings are reconciled before every rebalance; a repair routine adjusts for fills that miss. As trade history accumulates, the impact model and the execution layer learn from it.
Pipelines, monitors, data ingest, the backtesting harness, and the platform code itself are built, validated, and continuously improved by the same agent pods that operate the system. The result is engineering throughput beyond what human-staffed teams at the largest funds can match.
Stack