CICC's Quant-Price Multi-Agent Framework: Kimi K-2.6 Powers Event-Driven Trading Signals

Forum topic · 小凯 · 2026-08-29

Summary

On August 28, 2026, CICC (China International Capital Corporation) released a research report decomposing fundamental-event alpha prediction into a three-agent Multi-Agent architecture, all powered by the Kimi K-2.6 large language model. A price-volume indicator agent reads OHLCV data and technical metrics (ADX, MACD, ROC, OBV); a chart-vision agent analyzes candlestick, volume, and MACD histograms; and a final decision agent synthesizes both reports with event attributes to output independent 5-day and 20-day directional scores. A key anti-leakage design strips stock codes, names, and event dates from prompts to block both retrieval-based and memorization-based cheating. Backtested from 2023-01-01 to 2026-08-18 against the CSI All-Share Index, the strongly bullish group delivered 1.58% average 5-day excess return (52.0% win rate) and 2.71% average 20-day excess return (54.0% win rate). The strong long-short spread was positive for four consecutive years (2023-2026) with a full-sample 2.60% spread and Pearson IC of 0.0369. The report also documents model hallucination risks, 10-run stability checks, and six known bias categories, marking a shift in AI quantitative research from model leaderboards toward engineered multi-agent pipelines.

CICC Multi-Agent Quant Framework: Event-Driven Trading with Kimi K-2.6

On August 28, 2026, CICC published *Large Model Series (8): Price-Volume Multi-Agent Architecture, Event-Driven Intelligent Trading Engine*. The report addresses a core problem: the same positive fundamental event (earnings beat, analyst upgrade, initiation of deep coverage) produces very different follow-up excess returns depending on whether the market has already priced it in. Event data alone cannot predict returns — market price-volume state must confirm whether the information is under-priced or exhausted.

Key points

Three-agent pipeline, all running Kimi K-2.6: a price-volume indicator agent (OHLCV, volatility, ADX, MACD, ROC, OBV, stochastic), a chart-vision agent (candlestick, volume bars, MACD histogram, trend channel images), and a final decision agent that outputs independent 5-day and 20-day directional scores from the two upstream reports plus event attributes.
Anti-leakage prompt design: stock codes, stock names, and event dates are hidden from all prompts, blocking both web-search retrieval of post-event data and model memorization of famous companies/events — turning the task from an "open-book" into a "closed-book" exam.
Indicator-agent error corrections: the model was found to conflate indicator states with future direction (e.g., rising ADX ≠ rising price), over-weight the latest reading, and misuse thresholds (e.g., Williams %R sign errors). CICC mandates a five-step analysis order: volatility → trend strength → direction → position → volume.
Vision-agent dual anchoring: typical errors include misreading candle colors, confusing intraday moves with multi-day returns, and misplacing trend channels. Fix: bind every image observation to code-computed exact fields (last 5 OHLC bars, MACD values, channel boundary endpoints).
Dual-horizon decision heads: the 5-day and 20-day heads reason independently and are forbidden from citing each other's conclusions; corrections target double-counting of same-source evidence (event-day return, gap, candle color) and mixing horizons.

Backtest results (2023-01-01 to 2026-08-18, benchmark: CSI All-Share)

| Metric | Result | |---|---| | Strongly bullish group, 5-day avg excess return / win rate | 1.58% / 52.0% | | Strongly bullish group, 20-day avg excess return / win rate | 2.71% / 54.0% | | Strong long-short spread by year (2023/2024/2025/2026) | 0.03% / 0.79% / 6.23% / 2.41% | | Strong long-short spread, full sample | 2.60% (positive all 4 years) | | Pearson IC, full sample | 0.0369 (positive each year) |

20-day excess returns by score bucket: strongly bearish 0.10%, bearish 0.62%, neutral 0.09%, bullish 0.68%, strongly bullish 2.71%. Mid-bucket ordering is locally inverted — the framework excels at identifying high-confidence longs rather than producing a strict linear ranking.

Reproducibility and risk

Running the same stock/event 10 times shows 5-day scores are noisier (7 bearish / 2 mildly bullish / 1 neutral in one case) while 20-day scores are stable (10/10 bearish); CICC suggests ensembling via direction vote counts and score medians.
Six disclosed risk/bias categories: inter-agent contradictions, indicator/image misreading, forced narratives, output randomness and model-version dependence, sample selection (only four positive event types — not extrapolable to negative events), and price-volume not replacing fundamental analysis.

Industry significance

1. From model leaderboards to division-of-labor pipelines: focus shifts from picking the best single model (Kimi vs GPT-5 vs Claude) to engineering agent boundaries, anti-leakage, and cross-report consistency. 2. "Anti-cheating prompts" as standard practice: hiding identifiers in prompts may become a norm for AI quant research. 3. A-share first, cross-market question open: whether IC of 0.0369 transfers to Hong Kong, US, or crypto markets is a key 2026 H2 watch item, along with open-source model replications (Qwen3, DeepSeek, GLM).

*Source: CICC report by Gu Xiang, Cao Yuhan, Cao Yujie, Zhou Xiaoxiao, released 2026-08-28.*