CICC's LLM agent loop that searched China A-share price-volume factors (report of 2026-08-03)
The subject is the reported evidence for the discovered factors, not the engineering of the loop or the research team.
Subject
On 2026-08-03 CICC's research department (China International Capital Corporation) published Large-model series (7): an automated factor-discovery engine based on Loop Engineering, a system that searches for daily price-volume factors in China A-shares without human input. Our source is the summary CICC's own research account reposted on Sina Finance on 2026-08-05 (snapshot kept, md5 2d05a20b…); we have not read the full PDF.
The system runs as an agent loop of the kind Claude Code popularised: every five minutes a round generates 100 candidate factor expressions (25% mutation, 25% crossover, 15% parameter perturbation, 15% random, 20% written by an LLM from a list of 13 "mechanism families"), sub-agents and rules screen them, and a hard-coded backtest scores the survivors against 11 filters. A checkpoint file keeps "the hash set of factors already tested". The report's headline: after 581 rounds, 16,939 candidates tested, 69 kept (0.41%); the top five combined have an excess Sharpe of 3.14 and 18.3% annual excess return.
Three inputs
| Input | Value |
|---|---|
| Knowledge cutoff | The LLM behind the 20% of LLM-written candidates is not named, so its cutoff is unknown. The scored window (2025-01-01 to 2026-07-08) is recent enough that a model released by mid-2026 may have read about which A-share patterns worked in 2025; that cannot be ruled out. |
| Window | Factor values on the last 600 trading days (early 2024 to July 2026). Every performance figure is stated for 2025-01-01 to 2026-07-08, 553 days, 1.51 years. |
| Attempts | 16,939 candidates fully backtested; 58,100 generated (581 rounds × 100) before screening. |
Finding
No discovered factor clears the bar for the search that found it, and none of the reported performance is out of sample. The window used to select the factors is the window used to report them: the filters require positive excess return in 2025 and in 2026 and a 2026 Sharpe above 0.5, and the results are then shown for 2025 and 2026.
- Method: t = Sharpe × √years on the report's own excess Sharpe ratios and its stated window (1.51 years). The bar for the best of N is Bonferroni, one-sided, as everywhere on this site.
- The system already records its own trial count. It does not use it: no figure in the report is adjusted for the 16,939 tries.
| What is reported | Excess Sharpe | t over 1.51 years |
|---|---|---|
| Top five combined (picked by Sharpe from the kept set; four flipped in sign) | 3.14 | 3.86 |
| Best single factor ("gap-trend divergence") | 2.74 | 3.37 |
| "Amplitude expansion and gap premium" | 2.64 | 3.25 |
| Average of the class representatives | 1.63 | 2.01 |
| Average of all 69 kept | 1.60 | 1.97 |
| Average of factors found after the dominant structure was banned | 0.92 | 1.13 |
| If the search counted as | Bar | Clears it |
|---|---|---|
| 16,939 backtested | 4.69 | none |
| 58,100 generated | 4.94 | none |
| 69 kept, the rejected ignored | 3.41 | only the top-five blend, which was itself picked from the 69 |
At the search's real size a factor would need an excess Sharpe of about 3.8 over this window. The candidates are far from independent (the report says the overnight-gap field appears in 85% of the kept factors, and 51 of the 69 were found in one 50-round burst), so the effective number of tries is smaller than 16,939 and the Bonferroni bar is too strict. But it is not 69: the rejected 16,870 were rejected by looking at the same 553 days.
What the report says itself
- Its risk note: tested on history, "cannot ensure out-of-sample performance".
- It credits the 11 filters with avoiding the risk of the metric being fitted back (Goodhart's law). Filters scored on the reporting window are part of the fit, not a guard against it.
- It reports concentration honestly: overnight-gap factors dominate because they did best "in the 2025–2026 A-share environment", and the factors found later average a Sharpe of 0.92 against 1.68 for the core family.
- Two counts disagree inside the summary: the representatives are "21 classes" in one paragraph and "28" in the conclusion, with the same average Sharpe of 1.63.
Why it matters beyond one report
This is the case the time and trial axes of this site were built for, at an institution's scale: an LLM agent loop that logs every candidate it tests, then reports the best of them without the count. Generating factors has become nearly free; the 16,939 tries took about three days. What did not get cheaper is the out-of-sample record that tells the best of them apart from the luckiest.
Limits
- We read CICC's own summary, not the full report; the full PDF may contain a holdout we have not seen.
- Sharpe ratios are the report's, quoted, not recomputed; the factor expressions and return series are not public.
- The excess return is long top decile against the equal-weighted market with 0.1% one-way cost, as stated; we have not checked capacity or turnover.
- Converting Sharpe to t this way assumes weekly returns that are roughly independent; overlapping 5-day holdings would make the true t lower, not higher.
What would fix it
Publish the kept factors' returns from 2026-07-09 onward, a window none of them was selected on, and report every Sharpe next to the number of candidates tested to find it. The loop already stores that number.
Recheck: when CICC publishes a follow-up with out-of-sample results, or after 2026-12-31 against any forward figures released by then.