StockASkill puts stock-market calculations on a local database for AI agents

Keeping the calculations close to the data

StockASkill’s central design choice is to run analysis against a locally stored database rather than repeatedly request information from online financial services. The authors describe an embedded DuckDB database containing about 42.44 million daily price records and 186,432 quarterly financial records, spanning 26 years across mainland Chinese, Hong Kong and US equities, plus exchange-traded funds.

Online services still supply updates. The system checks the latest stored dates, downloads missing records and overlaps updates by three trading days to capture recent corrections. Alternative data sources and temporary backoffs are intended to handle connection failures. Local-first therefore means analysis does not depend on an immediate network response—not that the system never needs the internet.

The intended division of labour is straightforward: a language-model agent develops qualitative hypotheses and sets goals; StockASkill performs screening, allocation and risk calculations. Its machine-readable mode sends JSON results directly through standard output, with diagnostic messages kept separate. The paper’s “zero-disk” description concerns this exchange of results, not the underlying database storage.

What the speed tests show

On an AMD Ryzen 9 7950X workstation with 64 GB of memory, the authors report screening 5,421 Chinese A-share stocks in 0.38 seconds. Their SQLite-and-Python-loop baseline took 28.45 seconds, while a multithreaded Pandas implementation took 14.12 seconds. The reported improvement over SQLite was approximately 75-fold.

All three screening benchmarks recorded no network traffic. That comparison therefore supports a computational speed advantage in the tested implementations, rather than measuring the benefit of removing remote requests. A separate single-stock diagnostic comparison took 0.08 seconds locally versus 4.82 seconds through a remote API. Portfolio optimization, including lot rounding, reportedly took 0.41 seconds.

These are measurements from the authors’ specified workstation and workloads, not universal performance guarantees.

From rankings to tradable quantities

The stock-selection model combines seven familiar investment signals—value, quality, growth, momentum, short-term reversal, low volatility and company size—with an eighth based on analyst forecasts. It limits extreme observations and standardizes measurements before combining scores.

An optimizer then seeks a favourable balance between expected returns and volatility, subject to return targets, position limits and drawdown constraints. Drawdown measures the fall from a portfolio’s previous peak. During optimization, the system uses a tail-loss measure as a smoother substitute for the drawdown constraint, then checks the resulting historical portfolio path directly.

The software also adjusts equity exposure according to market breadth: the proportion of stocks above specified moving averages. Its stated allocations are 100% equities in bullish conditions, 60% in neutral conditions and 25% in defensive conditions, with the remainder held in cash.

For the Chinese A-share execution case, allocations are converted into 100-share lots while preserving a non-negative cash balance. A separate tracking system records investment rationales and issues exit tickets when trailing-stop thresholds are breached. Issuing a ticket, however, is not evidence that an order can always execute.

Strong simulations, with important boundaries

The reported investment tests cover January 2018 through October 2024. They simulate a 500,000-renminbi account, select ten stocks from historical CSI 300 and CSI 500 constituents, and rebalance every 21 trading days. Assumptions include 15 basis points of round-trip trading friction and an 8% trailing stop.

The momentum-enhanced configuration reportedly returned 18.42% annually, with an 11.85% maximum drawdown. The low-volatility configuration returned 12.65%, with a 5.82% drawdown. The CSI 300 buy-and-hold benchmark returned 3.12%, with a 39.54% drawdown. Reported Sharpe ratios, which express returns above a risk-free rate relative to volatility, ranged from 1.48 to 1.62 for the three strategy configurations, versus 0.22 for the benchmark.

The authors attribute approximately 58% of drawdown reduction to shifting exposure into cash, rather than stock selection alone. This matters because the strategies’ changing market exposure differs substantially from the fully invested benchmark.

The manuscript says historical constituent lists and records of subsequently delisted securities address survivorship bias. Nevertheless, it does not clearly specify a separate held-out testing protocol, and the reported investment results concern Chinese index constituents—not every market supported by the database. Historical constraint checks cannot guarantee future drawdown limits, particularly during trading halts or abrupt market changes.

The preprint presents a software architecture and encouraging author-reported simulations, not demonstrated live-trading performance. Multi-asset risk parity, learned nonlinear signals and adversarial multi-agent arrangements remain proposed extensions. The authors describe the software as research and educational material, not investment advice.