acceptodds
Under review as a conference paper at ICLR 2027

Risk-Aware Stock Recommendation via Functional Causal Modeling

Abstract

Stock recommendation aims to rank stocks by their expected returns based on historical market data. Most existing methods model inter-stock correlations, which may fail to capture the directed information flow induced by asymmetric relations such as supplier-consumer dependencies. They also often prioritize profitability without explicitly accounting for stock-specific risk. To address these limitations, we propose DISCO, a unified risk-aware stock recommendation framework that combines temporal causal discovery with functional causal modeling. DISCO learns directed relations among stocks and models dynamic risk as a stochastic noise variable, enabling it to jointly estimate future returns and stock-level uncertainty. An optional multi-situation inference strategy further accounts for uncertainty in both the learned causal graph and stock risk. Experiments on three real-world stock markets demonstrate that DISCO consistently outperforms strong baselines in profitability and risk-adjusted performance. Code is available at https://anonymous.4open.science/r/DISCO-D088.

Primary area:
learning on time series and dynamical systems
Keywords:
stock recommendation, causal discovery, functional causal models, dynamic risk modeling, quantitative finance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1RpSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposed the stock recommendation framework DISCO (integrating dynamic risk modeling and casual discovery) to rank stocks by predicted returns. DISCO learns a distribution over lagged, directed stock graphs by training the probability of edge existence jointly with the prediction model using a sparse graph prior and a variation objective. Then a functional causal model then aggregates information only from the selected parent stocks, weights it by learned relationship strengths, and produces a market-driven return estimate for each stock. DISCO adds a stock-specific noise term drawn from a skew-normal distribution to model the remaining return uncertainties. An optional inference procedure averages predictions from multiple graph and noise samples. On NASDAQ, NYSE, and TSE datasets, DISCO is reported higher Sharpe ratios and investment returns and than the compared baselines.

Critical strengths
  1. DISCO first implies graph casual discovery and functional casual model on a stock recommendation ranking question.

  2. DISCO reveals a significant improvement compared to the baseline methods, including newly proposed ones.

Critical weaknesses

The main concern is that the central architecture of DISCO closely resembles the work CausalStock [1], including formulae in the over lagged graphs casual discovery and in the FCM portions. See 4.3, 4.4 in [1]. Predicting continuous returns and optimizing stock rankings instead of classifying price movements is a meaningful adaptation, and some technical skills are adopted in this paper to modify the methods in [1], such as directly training the edge-probability parameters in graph casual discovery and selecting a skew-normal distribution instead of normal Gaussian distribution to model the noise. However, how much the reported improvement depend on these and how much comes from the inherited causal architecture in [1] is not clear. In particular, replacing the learned graph with a fully connected graph in ablation cannot answer that question. A more informative comparison would adapt [1] to the same price inputs, return target, and ranking objective, then separately test the effects of the skew-normal noise model and multi-situation inference. Without these controls, the necessity of DISCO’s added components remains unclear.

[1] Li, S., Sun, Y., Lin, Y., Gao, X., Shang, S., & Yan, R. (2024). CausalStock: Deep end-to-end causal discovery for news-driven multi-stock movement prediction. Advances in neural information processing systems, 37, 47432-47454.

Questions for authors
  1. See the weakness. The similarity between DISCO and [1] makes it difficult to assess the independent contribution of this paper. Could the authors explain precisely how DISCO differs from [1] and clarify how much each new component contributes to the reported gains?

  2. At inference, DISCO ranks stocks by the average return obtained from graph and risk samples. Could the ranking also reflect a investor’s risk preference, for example, by penalizing volatility or downside risk measured by the Value-at-Risk or conditional Value-at-Risk of the returns? Despite this may reduce the performance on indexes such as IRR.

Reviewer_WZv9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DISCO learns a distribution over lagged stock relations and uses an additive-noise model to predict returns. Stock-specific noise follows a learned skew-normal distribution, and an optional inference procedure averages predictions over sampled graphs and noise.

Critical strengths

Learning stock relations and return distributions together gives a useful way to study how uncertain relations affect stock ranking.

Critical weaknesses
  1. The noise parameters in Eq. (5) are static for each stock, and Section 5.2 ranks sample means without a stated risk-sensitive decision rule, leaving the dynamic risk claim unsupported.
  2. Masking the LSTM outputs in Eq. (4) does not guarantee the claimed exclusions of raw lagged inputs, since a retained hidden state can already contain information from a masked lag.
Questions for authors
  1. Which noise parameters change with the date or market history, and how does predicted uncertainty affect selection beyond estimating expected returns?
  2. How does the masking in Eq. (4) ensure a zero derivative with respect to an excluded raw lag after the LSTM has mixed observations across time?
  3. Can the authors show some graph-recovery evidence, or compare against a learned sparse predictive graph, to support reading the discovered links as causal?
Reviewer_e3GKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DISCO, a risk-aware stock recommendation framework that combines temporal causal discovery with functional causal modeling (FCM). DISCO learns a posterior distribution over directed relations between historical stock states and current returns, then models each stock’s return as the sum of a market-driven component and a stochastic noise component representing stock-specific risk. It also introduces an optional multi-situation inference to reduce uncertainty. Experiments on three stock-market datasets show that DISCO achieves the best results among the diverse baselines on both metrics: IRR, and SR.

Critical strengths

DISCO’s main strength is its attempt to model directed temporal causal dependencies among stocks and incorporate them into stock recommendation, rather than relying on inter-stock undirected correlations. It also explicitly models stock-specific risk through a stochastic component. These two aspects are integrated within a unified framework, and the experiments on three stock-market datasets show that DISCO achieves the best reported results among the evaluated baselines in both profitability and risk-adjusted performance. This provides empirical support for the value of jointly modeling inter-stock causal relations and risk. Furthermore, The optional multi-situation inference strategy is a useful extension of the framework. It samples from both the learned causal-graph posterior and the stock-specific risk distributions, then aggregates the resulting return predictions. This provides a principled way to account for uncertainty at inference time, and the reported results show improvements over standard inference.

Critical weaknesses

W1. Scalability of the causal graph modeling. DISCO parameterizes the temporal causal graph using edge-probability tensors whose size scales as O(LD2)O(LD^2), where DD is the number of stocks and LL is the number of time lags. It also introduces a separate causal-strength tensor of the same order. This quadratic dependence on the number of stocks may limit the method’s applicability to substantially larger stock markets. Although the paper reports DISCO’s computational complexity and inference time (at Table 3), it does not compare memory usage or inference time against the baselines. Such comparisons would help clarify the practical cost of the proposed graph modeling.

W2. Limited evaluation of the additive-noise FCM. The method relies on an additive-noise FCM, in which the return is decomposed into a market-driven component and a stochastic residual representing stock-specific risk. However, the ablations do not isolate the contribution of this FCM formulation itself by comparing it with alternative ways. They do not establish whether the additive-noise FCM is necessary or whether another formulation could better capture the information not explained by historical causal stock features.

Questions for authors

Q1. Could you compare DISCO’s memory usage and inference time with those of the baselines under matched settings?

Q2. Could you provide ablations that compare the additive-noise FCM with alternative ways of FCMs modeling the residual or conditional return distribution?

Q3. Could you evaluate the effect of the sparse graph prior by comparing it with a dense-graph variant or another setting without the sparsity constraint?

Q4. Could you evaluate DISCO on a larger stock-market dataset and report its computational efficiency in that setting?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.