- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes MEME, an LLM-based stock selection framework motivated by Narrative Economics. An LLM extracts bullish/bearish "investment arguments" for each stock from fundamental, news and technical data. These arguments are embedded and clustered daily with a GMM into "Modes of Thought". Modes are matched across days via Hungarian assignment, and each mode is scored by an EMA of its realised, polarity-signed excess returns. A stock's signal is the mode-score-weighted difference between its bullish and bearish arguments, and the portfolio holds the top-K stocks with equal weights. Experiments on SSE 50, CSI 300, CSI 500 and Nasdaq 100 (2023Q4–2025Q3, plus a 2025Q4 out-of-sample quarter) report gains over seven baselines, together with ablations and a case study linking modes to value and trend-following styles.
Critical strengths
-
The motivation behind is interesting. Clustering LLM-generated investment arguments and asking which types of reasoning are currently profitable is an appealing alternative to black-box return prediction. The case study gives some qualitative insight, for example dividend and balance-sheet arguments working in the bear market and momentum arguments in the rally.
-
The evaluation covers four stock pools in two markets and seven baselines from ML, DL and LLM-agent families. It also includes ablations, sensitivity analysis, a signal-orthogonality analysis, an earnings-season split and a post-cutoff test.
-
Transparency. Code and prompts are released. The limitations section openly reports a case where MEME underperforms (CSI 1000, where MASTER beats it).
Critical weaknesses
-
The method is naive. Plugging Eq. 12 into Eq. 13 is basically a linear factor model on K text clusters, where each factor's return is forecast by an EMA of its own past returns. With λ = 0.5 the memory is about one period, and since K is fixed with a forced one-to-one matching, modes can never "emerge or decay". The narrative-lifecycle framing therefore has no real counterpart in the method.
-
Data leakage is not tackled. The main period (2023Q4–2025Q3) largely falls within DeepSeek V3.1's training data, and the post-cutoff test covers only one quarter. In that quarter MEME's advantage largely disappears (CSI 300 IC 4.12 vs. 4.09), and its IC drops by 39%, while DTML and LightGBM barely change.
-
The backtest is unrealistic. The transaction cost of 1.5e-4 is below the A-share stamp duty alone (5e-4 on sells since Aug 2023). With daily rebalancing and no reported turnover, the AR and SR results are not credible.
-
The GMM is ill-posed. It uses full covariance on 4096d embeddings with only about 40–200 samples per component, so each $ \Sigma_k $ has rank at most and the remaining eigenvalues equal reg_covar = 1e-6. The "soft assignment" explanation for why PM beats k-means does not hold.
-
Minor. Missing references:
- Narrative Asset Pricing: Interpretable Systematic Risk Factors from News Text. RFS, 2023.
- Factor Momentum and the Momentum Factor. JF, 2022.
- The Memorization Problem: Can We Trust LLMs' Economic Forecasts? arXiv, 2025.
- Can LLM-based Financial Investing Strategies Outperform the Market in Long Run? KDD, 2026.
Questions for authors
Please see the weakness section