CICC's Price-Volume Multi-Agent Framework: Kimi K-2.6 Delivers 1.58% 5-Day / 2.71% 20-Day Excess Returns on Positive Fundamental Events

Forum topic · 小凯 · 2026-08-29

Summary

CICC (China International Capital Corporation) published a report on August 28, 2026 introducing a price-volume Multi-Agent architecture for event-driven trading, built entirely on the Kimi K-2.6 model. The framework splits the analysis of positive fundamental events (earnings beats, analyst upgrades, deep report coverage) across three agents: a quantitative indicator agent reading OHLCV, volatility, ADX and MACD; a chart-recognition agent analyzing candlestick, volume and MACD images; and a final decision agent that synthesizes both reports with event attributes to output independent 5-day and 20-day direction scores. To prevent information leakage, prompts hide stock codes, names, and event dates. Backtested from 2023-01-01 to 2026-08-18 against the CSI All Share Index, the strong-bullish group achieved 1.58% average 5-day excess return (52.0% win rate) and 2.71% 20-day excess return (54.0% win rate); the strong long-short spread was positive for four consecutive years (full-sample 2.60%), with a full-sample Pearson IC of 0.0369.

CICC's Price-Volume Multi-Agent Architecture: Event-Driven Intelligent Trading Engine (Kimi K-2.6)

On August 28, 2026, CICC released *Large Model Series (8): Price-Volume Multi-Agent Architecture, Event-Driven Intelligent Trading Engine*. The report decomposes the post-event excess return question for positive fundamental events (earnings beats, earnings forecast upgrades, deep analyst coverage) into three specialized agents — all powered by Kimi K-2.6.

Framework Overview

Quantitative indicator agent: reads OHLCV, volatility, ADX, MACD, ROC, OBV, stochastic and volume ratio data; outputs a standardized market-state assessment with a directional view.
Chart-recognition agent: receives candlestick, volume-bar and MACD histogram images plus key numeric fields; assesses overall price structure and 5/20-day direction.
Final decision agent: combines event attributes with the two upstream reports, outputting independent 5-day and 20-day direction scores plus a risk-reward ratio. The two horizons are required to reason independently without referencing each other.

Anti-Leakage Prompt Design

A key engineering choice: prompts hide the stock code, stock name, and event date, blocking both (1) real-time retrieval of post-event price action and commentary, and (2) pretraining memory of well-known companies and historical episodes. Agents can only judge from information available at the analysis point.

Correcting Known LLM Biases

The report documents and mitigates systematic failure modes:

Indicator agent: conflating indicator states with future direction (e.g., rising ADX ≠ up move); over-weighting the latest reading; mechanically applying thresholds (e.g., Williams %R sign errors). Mitigation: a mandatory five-step analysis order — volatility → trend strength → direction → position → volume.
Vision agent: misreading candlestick color, confusing intraday direction with cross-day returns, misplacing trendlines/channel boundaries, threshold errors. Mitigation: "dual anchoring" — image intuition cross-checked against code-computed exact fields (last 5 OHLC bars, 5 MACD histogram values, channel endpoints).
Decision agent: double-counting same-source evidence (event-day return, gap, candle color, close position counted as one evidence set); interpreting market state as direction; mixing horizon-specific evidence. Mitigation: symmetric scoring thresholds, shrinkage toward neutral when evidence conflicts.

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

| Metric | Result | |---|---| | Strong-bullish group, 5-day avg excess return | 1.58% (win rate 52.0%, ranked 1st of 5 groups) | | Strong-bullish group, 20-day avg excess return | 2.71% (win rate 54.0%, ranked 1st) | | Strong long-short spread, 2023–2026 | Positive all 4 years: 0.03% / 0.79% / 6.23% / 2.41%; full sample 2.60% | | Pearson IC (20-day excess) | Full sample 0.0369 (positive every year; 2025 peak 0.0904) |

Middle buckets (mildly bearish vs neutral) show local inversions, indicating the framework is best at identifying high-confidence opportunities rather than strictly linearly ranking all samples.

Reproducibility

Running 10 independent trials on the same stock-event shows 5-day scores fluctuate (7 bearish / 2 mildly bullish / 1 neutral in one example), while 20-day scores were consistently bearish (9 of 10 strong-bearish). The authors recommend ensemble judgment via direction vote share, score median, and dispersion.

Stated Risks

Six acknowledged biases: cross-agent report contradictions; indicator/image misreading; forced narrative application; output randomness and model-version dependence; sample selection (only four positive event types — not extensible to negative events); and price-volume signals not substituting for fundamental research. Recommended to combine with fundamental analysis, portfolio constraints, and risk budgets.

Significance

The report marks a shift in AI quant methodology: from single-model benchmark competition to multi-agent division-of-labor pipelines, and from "what to ask" in prompt engineering to "what to hide" — treating anti-leakage prompts as a closed-book exam for AI quant research.

Things to Watch (H2 2026)

1. Cross-market transfer of the framework (HK, US, crypto) given the 0.0369 A-share IC; 2. Whether anti-cheating prompt design becomes an industry standard; 3. Whether open-source models (Qwen3 / DeepSeek / GLM) can match Kimi K-2.6 in this pipeline.

References

1. CICC, *Large Model Series (8): Price-Volume Multi-Agent Architecture, Event-Driven Intelligent Trading Engine* (2026-08-28; authors: Gu Xiang, Cao Yuhan, Cao Yujie, Zhou Xiaoxiao) 2. Kimi K-2.6 (Moonshot AI, spring 2026 version) 3. CSI All Share Index; Wind database 4. CICC fundamental quant series (23): Ten Q&As on A-share event impact (prior research)