acceptodds
Under review as a conference paper at ICLR 2027

LiFT: Neurosymbolic Conditioning for Portfolio Allocation

Abstract

Deep learning portfolio allocators achieve strong in-sample performance but degrade in out-of-sample performance during regime shifts. To mitigate this, recent deep portfolio allocators add market knowledge and investment rules at three levels: the input, where regime labels are added to the features; the representation, where a conditioning signal modulates each stock's encoding; or the output, as constrained decision layers or differentiable logic penalties on portfolio limits. Each is demonstrated on a single architecture, so the levels cannot be compared, and no gain can be attributed to the underlying knowledge. Supplying all three to any encoder raises three challenges. A regime describes the current market state, not a stock, so it has no natural place in a stock's input series. Trading rules are discrete, giving no usable gradient, and their relative weights are undetermined. Risk rules cap holdings but never require a minimum, so holding only cash satisfies them all. To address all three, we present , a neurosymbolic framework that operates at the input, representation, and output level of any time-series encoder. At input, market-state features join the per-stock inputs, At the representation level, binary regime indicators modulate each stock's encoding via FiLM. At output, trading rules become differentiable penalties, with each rule's weight learned by dual ascent. A minimum-investment rule and a cash ceiling prevent the all-cash solution. Being encoder-agnostic, the three additions run unchanged on five encoders, two stock universes, and four walk-forward windows at matched capacity. Together they condition allocation on market state and penalize rule violations during training. LiFT improves out-of-sample Sharpe over the baseline by 50% on the DJIA, on all five encoders, and by 25% on the S&P100, with Sortino and Calmar also improving. Separating the levels reveals which knowledge drives these gains.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
neurosymbolic conditioning, differentiable logic, feature-wise linear modulation, representation learning, time series transformers, domain knowledge injection, portfolio allocation, finance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_o12POpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LiFT, a modular framework for adding market knowledge to neural portfolio allocation at three levels. At the input level, four market-state features are appended to each stock’s features. At the representation level, binary regime indicators modulate stock representations using FiLM. At the output level, seven portfolio rules are converted into differentiable penalties whose weights are updated by dual ascent. The framework is evaluated using five time-series encoders, two US stock universes, four chronological test windows, and 25 random seeds, with transaction costs included. Full LiFT improves Sharpe over a price-only baseline on average. However, the ablations show that the input features are the only component that reliably improves both universes. Conditioning is neutral or harmful, full LiFT is worse than the input-only model on the S&P 100, and shuffling the regime indicators does not change performance. No learned method beats equal weighting under the primary daily-rebalancing setting.

Critical strengths
  1. Broad and careful empirical comparison. The paper evaluates five different encoders, all eight combinations of its three components, four chronological test windows, two stock universes, and 25 seeds. It includes transaction costs, several classical baselines, multiple portfolio metrics, and a holding-period analysis.
  2. Useful component analysis and transparent reporting. Separating the input features, conditioning, and rules reveals which components contribute to performance. The authors openly report negative findings, including the failure of conditioning, the shuffled-indicator result, the dependence on one encoder, and the inability of learned models to beat equal weighting under daily rebalancing.
Critical weaknesses
  1. The reported results do not provide convincing evidence for the proposed neurosymbolic and regime-conditioned components. The input-only model is the only configuration that reliably improves both universes. Full LiFT is worse than the input-only model on the S&P 100. Conditioning alone is neutral or harmful, and the incremental DJIA improvement is statistically clear for only one encoder. Most importantly, shuffling the regime indicators across time produces no detectable performance change. This suggests that the improvement does not depend on correctly identifying the current market regime.
  2. The statistical and backtesting design does not fully support the out-of-sample performance claims. The bootstrap resamples training runs, but the 25 seeds share the same four realized test periods and therefore are not independent samples of market performance. No time-series bootstrap over daily returns is reported. In addition, the historical universes are constructed using current index membership, introducing survivorship and look-ahead universe selection. These issues can affect relative method rankings, not only absolute performance levels.
  3. The rule mechanism is not sufficiently validated or isolated. The paper does not report out-of-sample rule satisfaction, compare learned rule weights with fixed weights or simple penalties, or cleanly separate stock-specific conditioning from the rule component. Since the rules are applied only during training, it is not clear whether LiFT actually produces more compliant portfolios at test time or whether its limited DJIA improvement could be obtained through simpler turnover and exposure regularization.
  4. The practical improvement is limited. No learned method beats equal weighting under daily rebalancing. With less frequent rebalancing, LiFT becomes competitive on the DJIA, but its advantage over the input-only model nearly disappears; on the S&P 100, it remains below several classical methods. This weakens the significance of the improvement over the internal price-only baseline.
Questions for authors

The main questions affecting my assessment are already described in the Critical Weaknesses. In particular, they concern whether full LiFT provides reliable improvement over the input-only model, whether the conclusions remain under time-series-aware uncertainty estimates and point-in-time index membership, and whether the rules improve out-of-sample rule satisfaction relative to simpler regularization. I have the following additional clarification questions:

  1. Appendices C and F contain bracketed placeholders instead of the referenced results.

  2. Are the weekly and monthly results obtained by applying models trained for daily allocation at lower rebalancing frequencies, or by retraining the models for each frequency? What rebalancing frequency do the authors consider the intended deployment setting?

Reviewer_YMf5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Logic-informed Feature-wise Linear Modulation Training (LiFT), a model-agnostic framework injecting domain knowledge into deep portfolio allocators at three levels: market-state features at the input, binary regime indicators via Feature-wise Linear Modulation (FiLM) at the representation level, and differentiable trading rules (weighted via dual ascent) in the training loss. Experiments were carried out across five encoders and two universes (the DJIA and SP100).

Critical strengths
  • The paper provides an approach to add market information, regime indicator, and trading rule to a deep learning based strategy. These additional info is presumed to be valuable to improve trading performance.
  • Experiments are conducted on a wide variety of encoder. Ablation experiments show good attribution.
Critical weaknesses
  • The method relies on known techniques and heavy feature engineering. Not sure how new the proposed method is from a machine learning perspective.
  • The regime indicator and the trading rules do not seem to help from the ablation. It appears that the market features do most of the heavy lifting.
Questions for authors
  • eq (4) l2 penalty is wrong.
  • minor typo line 024 comma should be period.: At input, market-state features join the per-stock inputs, At the representation level,
  • ablation design section is hard to parse.
Reviewer_YWzNOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes LiFT, framework for incorporating domain knowledge into neural portfolio allocation at three levels: input features, representation conditioning, output-level rules. The proposed method is evaluated across five time-series encoders, two stock universes

Critical strengths
  • The paper presents careful experimental design.

  • Figure 1 provides a clear overview of the method by representing the decomposition of how knowledge enters the model at the input, representation, and objective levels

Critical weaknesses
  • The paper is difficult to follow. The text is hard to parse. Most importantly, the paper might violate the double-blind review (line 487)

  • It's not clear the main attribution for model performance is neurosymbolic. The most robust improvement comes from simply appending market-state features. On S&P 100, LiFT(F) reaches a higher Sharpe than full LiFT, while adding conditioning to F decreases Sharpe substantially.

Questions for authors

See weakness.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their careful reading. The reviews raise three common points:

  1. The symbolic components, not the input features, carry the claimed gain.
  2. The uncertainty estimates and the universe construction.
  3. The rule mechanism is not isolated or validated out of sample.

We have revised the paper and added experiments the reviews ask for. The revised PDF contains the following changes.

  1. Corrections.
    • Line 487 named an author; the name is removed.
    • Equation (4) now uses the squared L2 norm, which the implementation always used.
    • The comma at line 024 is now a period, and several typographical errors are fixed.
    • Section 6.2 said the rules limit single-stock and sector weight; it now says their effects are small.
    • Appendix E said daily returns were kept for five seeds; they exist for all 25.
  2. Precise claims.
    • Against a price-only baseline, LiFT raises Sharpe by 50% on the DJIA (all five encoders) and 25% on the S&P 100 (four of five).
    • The market-state features are the only level that helps in both universes.
    • Conditioning alone gives no reliable gain, and its shared variant lowers the S&P 100 Sharpe ratio.
    • The rules act as regularization: their effects on compliance, idle cash, and turnover are all small.
    • The paper no longer claims that the symbolic levels are the source of the improvement.
  3. New experiments, each with the appendix that reports it.
    • Appendix E: a day interval, a block bootstrap over test days, for every Sharpe contrast and the headline Sortino and Calmar gains.
    • Appendix C: two further test windows, July 2015 to June 2019, with 500 runs per universe.
    • Appendix K (new): out-of-sample rule compliance from stored daily weights, in two batches and against a generic penalty.
    • Appendix L (new): the rules against the generic turnover and weight penalties at 100× their defaults.
    • Appendix F: learned against hand-set rule weights.
    • Appendices C and F, which contained bracketed placeholders, are complete.
  4. Limitations, stated in the main text.
    • The universes use current index membership, so survivorship enters when the data are fetched.
    • No learned allocator beats equal weighting under daily rebalancing.
    • On the S&P 100, the day interval for the headline gain reaches -0.003 at its lower bound.

The individual responses give the numbers for each point

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.