QuantG developer adds holding period to factor definitions

The three stages trace a shift in what a factor definition is for. In the first stage the developer ported the flow of RD-Agent, an open-source framework for automating factor research, and handed the LLM a daily table with 45 columns built from J-Quants data. The hypotheses that came back clustered on momentum, reversal and volatility, and the other columns went largely unused — not because the model lacked data, but because nobody had a way to constrain where it looked.

The second stage added a two-level research map of six themes and 19 definitions, each definition listing the columns the LLM was allowed to read, with one pool slot per definition so the pool would not tilt in one direction. Slow-disclosure definitions were re-mined roughly every 400 new disclosures while price definitions were re-mined every 3 business days, and the definition count passed 20. Yet in mid-October the pool was still almost entirely daily price and volume transforms, highly correlated with each other, so combining them simply selected high-volatility stocks.

The diagnosis pointed outside the definitions. When the developer re-scored pooled factors at 2, 5, 10 and 20 business days, turnover-liquidity factors rose from an IC of 0.032 to 0.051 at 20 days, while short-term return and market-breadth factors reversed direction at 20 days. A detour of extending every factor to 10 days looked good on 2026 data but collapsed to near zero out of sample on 2023–2024, which is why the current design ties k to the type of information a factor uses rather than to a single universal horizon. The third stage also narrowed each definition to one testable hypothesis and moved evaluation to buying at the next open and selling at the open k business days later.