Researchers face two core challenges when building quantitative strategies: discovering reliable alpha factors and turning those signals into robust, reproducible models. Traditional pipelines mix data preparation, hypothesis generation, backtesting, and model tuning in a tangled way, making it hard to audit each step, isolate failures, or reuse successful components. The result is wasted effort, opaque processes, and models that often overfit to noise.
AQuA solves this by splitting the workflow into two independent, agent‑driven loops that never talk directly to each other. In Part I, a six‑agent pipeline—Data Steward, Visual Analyst, Idea Miner, Factor Evaluator, Backtest Engineer, and Research Librarian—is coordinated by an AI Manager. Every handoff goes through the Manager, creating a clear, auditable trail. Factors start as falsifiable proposals (hypothesis, mechanism, predicted direction, refutation conditions) before being assembled from a standardized operator registry. Because each operator only reads a trailing time window or the current timestamp, causality is preserved when they are combined. Three feedback loops refine direction, update beliefs from falsification, and carry memory across runs to steer future searches.
Part II takes the discovered factors and builds predictive models with a config‑driven approach. Each experiment differs by exactly one configuration change—architecture, loss, sampler, or optimizer—so variants stay comparable. The model uses a multi‑scale 1‑D convolution front‑end, a configurable backbone (LSTM, Mamba, or attention), a cross‑sectional mixing stage, gated fusion, and a per‑stock readout. Training runs on 2010‑2019, with 2020 held out as an embargo gap and 2021‑2025 as untouched test data. The hybrid predictor achieves a per‑stock information coefficient of +0.0843, outperforming the best baseline (GRU at +0.0613) by 37.5 % relative gain.
When the signal is turned into a dollar‑neutral, sector‑neutral long‑short book with a 2‑bp transaction cost, the strategy delivers a held‑out Sharpe of +2.15. Adding a causal volatility‑targeting overlay lifts Sharpe to +2.50, and a fully walk‑forward, past‑data‑only parameter selection still yields +2.00. Returns are positive each year from 2021 through 2025, including the 2022 drawdown.
Key takeaways for practitioners:
– Two independent research loops eliminate shared state and simplify debugging.
– The agent‑managed DSL guarantees exploratory freedom while locking data, features, labels, and evaluation.
– Part I reaches ~0.190 combined IC on crypto; Part II delivers +0.0843 IC on US equities versus +0.0613 for a GRU baseline.
– The resulting equity book sustains a +2.50 Sharpe at 2 bps and remains profitable across all five test years.
#AI #QuantFinance #FactorDiscovery #ModelDevelopment #Automation #Research