acceptodds
Under review as a conference paper at ICLR 2027

Risk Analysis of Recurrent Neural Networks for Financial Time Series

Abstract

Forecasting financial time series is a fundamental task for modern quantitative finance. Recurrent neural networks (RNNs) have been widely applied in this domain, yet their statistical behavior under the structural features of financial data remains poorly understood. We study the excess risk of RNNs for price movement prediction with k-line chart data. To obtain a financially interpretable risk characterization, we introduce a tractable data-generating model and decompose excess risk into approximation and estimation errors. The approximation analysis separates finite temporal representation, nonlinear function approximation, and normalization-induced information loss. Under binary cross-entropy loss, we prove that the finite-width term decays as . We further show that normalized observations inherit the -mixing property of the underlying dynamics, yielding an estimation bound based on Rademacher complexity for dependent data. Combining the two error bounds, we reveal how financially interpretable factor into risk: the signal-to-noise ratio (SNR) proxy governs approximation difficulty, while the persistence parameter \(\rho\) controls effective statistical information. Notably, a larger SNR proxy can increase approximation error and hence excess risk. We conduct controlled experiments on simulated data. The results exhibit trends consistent with the theoretical analysis, and empirically reveal the approximation–estimation tradeoff across model capacity and sample size.

Primary area:
learning theory
Keywords:
recurrent neural networks, excess risk bounds, approximation theory, Rademacher complexity, weak dependent data, financial time series

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_2RY5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the statistical risks of recurrent neural networks (RNNs) for financial time-series prediction. It derives theoretical bounds on approximation and estimation errors and examines the effects of network width, historical window length, sample size, signal-to-noise ratio, and temporal dependence through synthetic experiments.

Critical strengths
  1. The goal of tying RNN risk bounds to financially interpretable quantities is worthwhile and more application-oriented than generic RNN bounds.
  2. It provides a theoretical analysis of both approximation and estimation errors.
  3. The experiments illustrate several trends consistent with the theoretical analysis.
Critical weaknesses
  1. The paper does not sufficiently distinguish its theoretical contributions from existing RNN approximation and generalization results.
  2. The VAR-based synthetic data generation framework may not adequately capture important financial characteristics, such as heavy-tailed returns, volatility clustering, and regime changes.
  3. The experiments show only qualitative trends and do not validate the tightness or practical relevance of the bounds, which lie orders of magnitude above the measured excess risk. Results are averaged over only five datasets with no confidence intervals, and more diverse data-generating processes would strengthen the claims.
  4. The paper links excess risk to financially motivated quantities (the SNR proxy and persistence), but the risk it analyzes is statistical classification risk, not financial investment risk. Since the label records only the direction of the price move and not its magnitude, it is unclear how the bounds relate to drawdown, tail risk, or risk-adjusted returns. The SNR proxy is model-dependent, and all conclusions are validated only on synthetic data. The authors should clarify the scope of their interpretability claims and discuss implications for real-world forecasting.
  5. Some proofs, particularly those associated with Lemmas 4.10 and 4.11, appear incomplete in the appendix. The derivation of generalization bounds under dependent observations also requires clearer justification.
Questions for authors

Please see the Weaknesses.

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

This paper studies the excess risk of RNNs, for a specific finance setting. The setting is predicting the direction of the next period price movements from OHLC observations. This paper uses its own data generation model in which the open and close dynamics follows a finite order vector autoregression, while intra period high and lows are generated through a Brownian bridge conditional on the end points.

In this paper, the excess risk is decomposed into approximation error and estimation error, where several statistical learning machinery were applied.

the authors have conducted experiments on simulated data to examine whether trained RNNs exhibit the trends suggested by the analysis, including the approximation estimation tradeoff.

Critical strengths

This work connects RNN's excess risk to the properties of the author's assumed data generation process.

On the theory side: good analytical efforts have made to the bound: this paper develops bounds for both the approximation error and estimation error. Many statistical machine learning tools were employed for this process. It also accounts for the overlapping input window and beta mixing observations.

on the experimental side: several controlled empirical studies were carried out on the simulated data with varying network width, available history length, predicability and sample size. It also checks whether the observed increase in the excess risk with predictability also holds when using a variance based SNR measure.

Critical weaknesses

A major drawback is that this paper's theory and experiments seems to be very restricted, and only tied to the data generation process described in the paper: The open and close dynamics follows a finite order vector autoregression, while intra period high and lows are generated through a Brownian bridge conditional on the end points. Therefore, my major concern, is whether such model was accurately represented in the real world to make it useful in finance. Conversely, what other non-financial time series this work can also apply to?

In B.2 in page 15, under the heading: B.2 PROOF OF LEMMA 4.10, the author misses the proof for this lemma.

Questions for authors

n/a

Reviewer_TJ1hOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the excess risk of recurrent neural networks for binary price-direction prediction using OHLC time-series data. The authors propose a stylized data-generating process in which intraperiod high and low prices are generated using a Brownian bridge, while price increments follow a finite-order vector autoregression. Their main contribution is a theoretical analysis that decomposes excess risk into approximation and estimation errors, accounting for information loss due to window-wise normalization and temporal dependence through β-mixing. The analysis characterizes how network width, signal-to-noise ratio, and temporal persistence affect the resulting risk bounds. The authors validate the theoretical findings through controlled experiments on simulated data, examining the effects of network width, input history length, predictability, and sample size.

Critical strengths

1: The paper analyzes RNN-based price-direction classification under a stylized OHLC data-generating process, explicitly accounting for window-wise normalization and temporal dependence. By incorporating these features into an approximation–estimation analysis, it connects RNN learning theory to structural properties of financial time series rather than treating observations as generic sequential data. In particular, the β-mixing inheritance result provides a useful link between the dependence structure of the underlying process and the statistical analysis of normalized OHLC observations. This domain-grounded formulation is a central strength.

2: The paper separates approximation error due to finite temporal representation, nonlinear function approximation, and normalization-induced information loss, while accounting for temporal dependence in the estimation analysis. It also identifies how an SNR proxy and persistence parameter enter the bounds. Notably, a larger SNR proxy can increase the approximation-error upper bound under window-wise normalization, highlighting that discarding scale information can complicate learning even when the underlying signal is more pronounced. This finding, together with controlled simulations examining the predicted qualitative trends, gives the theoretical analysis useful explanatory value, while its implications remain bounded by the stylized data-generating assumptions.

Critical weaknesses

1: The paper frames the problem in a real financial forecasting context, but all experiments use simulated data generated from a stylized VAR–Brownian-bridge process. While this controlled setting facilitates theoretical analysis (a scope the authors explicitly acknowledge in the conclusion), it does not establish whether the results extend to real financial time series exhibiting features such as volatility clustering, heavy tails, and regime shifts. In particular, the central finding that a larger SNR proxy can increase the approximation-error bound under min–max normalization may depend on the assumed signal structure and normalization scheme. Without real-data validation or experiments under richer data-generating processes, the generality and practical relevance of this finding remain unclear.

2: The experiments report qualitative trends consistent with the theoretical analysis, but such agreement does not establish that the derived bounds accurately characterize the magnitude of excess risk or provide actionable guidance for model selection. In particular, the paper does not systematically evaluate how closely the theoretical bounds track empirical excess risk across experimental settings, and the qualitative agreement does not establish quantitative predictive accuracy of the bounds for model capacity or history length. This leaves an important gap between establishing theoretically interpretable relationships and demonstrating that the resulting guarantees offer reliable quantitative insight into the behavior of trained RNNs.

Questions for authors

1: The paper intentionally restricts all experiments to the stylized VAR-Brownian-bridge DGP, as the authors acknowledge in the conclusion. Given that the paper’s motivation is grounded in real financial forecasting, could the authors include a preliminary validation on real OHLC data to examine whether the qualitative trends-particularly the SNR–approximation relationship-transfer to real markets? If this was attempted and obstacles arose, what were they?

2: The proof of Lemma 4.10 is referenced as Appendix B.2, but the provided appendix appears to contain only the section heading, with no proof text. Could the authors provide the complete proof or clarify where it can be found? This is necessary to assess the validity of the estimation-bound result.

3: The finding that a larger SNR proxy can increase the approximation-error bound is derived under window-wise min–max normalization. Is this relationship robust to alternative normalization schemes, such as z-score normalization, or does it depend on the specific information discarded by min–max preprocessing? Clarifying the scope of this result would help assess the generality of the paper's central theoretical insight.

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

The paper studies the excess risk of ReLU RNNs trained to predict next-period price direction from k-line (OHLC) histories. It introduces a stylized VAR-Brownian Bridge (VAR-BB) data-generating process and decomposes excess risk into approximation error and estimation error. The theoretical findings are evaluated through controlled experiments on synthetic OHLC data generated from the proposed VAR-BB model. The experiments investigate the effects of network width, input history length, predictability, and sample size. The results qualitatively support the predicted approximation trends and the approximation–estimation tradeoff.

Critical strengths

1.Structured risk decomposition. The paper decomposes the excess risk of RNNs into approximation and estimation errors. This provides a relatively systematic framework for studying how model capacity and preprocessing affect prediction risk. 2.Consideration of temporal dependence. The paper incorporates β-mixing into the estimation analysis and derives a Rademacher complexity-based bound for dependent observations.

Critical weaknesses

1.The theoretical analysis assumes a VAR(6)\mathrm{VAR}(6) process, but Algorithm 1 does not explicitly implement the VAR dynamics. This raises concerns about whether the experiments satisfy the theoretical assumptions, weakening the empirical validation.

2.All experiments use synthetic OHLC data. The analysis also focuses exclusively on min-max normalization and omits volume. It remains unclear whether the findings generalize to real financial data and alternative preprocessing schemes.

Questions for authors

1.Algorithm 1 does not appear to implement the VAR(6)\mathrm{VAR}(6) dynamics in Eq. (1): neither the lag matrices Φi\Phi_i nor BB appear, and the open and close increments look i.i.d. If the VAR enters through the time-indexed μDt\mu_D^t and σDt\sigma_D^t, please state how they are computed. The notation is also inconsistent (μD\mu_D vs. μDt\mu_D^t, σN\sigma_N vs. σNt\sigma_N^t).

2.The persistence parameter ρ\rho is presented as one of the two key financial quantities, but it is never varied in the experiments (Figs. 1–4 vary only mm, τ\tau, SNR, and nn). How does the empirical estimation error change with ρ\rho at fixed nn? Since changing the VAR persistence typically also changes Var⁡(st)\operatorname{Var}(s_t), how would the effect of ρ\rho be separated from that of SNR?

3.All experiments use simulated data, and no result on real candlestick data is given. Volume is omitted from the Kline inputs. "Volume provides information on information quality that cannot be deduced from the price statistic."[1].

4.The information loss arises from the choice of window-wise min-max normalization. Do the conclusions remain valid under other commonly used preprocessing schemes, such as window-wise zz-score normalization or return-based inputs (e.g., rt=log⁡(Pt/Pt−1)r_t = \log(P_t/P_{t-1}))?

[1]. Blume, Lawrence, David Easley, and Maureen O'hara. "Market statistics and technical analysis: The role of volume." The journal of finance 49.1 (1994): 153-181.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.