acceptodds
Under review as a conference paper at ICLR 2027

FiRE: Scaling and Aligning Financial Foundation Models

36 views

Abstract

We introduce FiRE, a financial foundation model that combines generative pretraining with reward-guided post-training. Generative pretraining offers supervision directly from historical market trajectories, but scaling this paradigm exposes two limitations: high-capacity token spaces make categorical next-token prediction increasingly expensive, while sequence-modeling objectives are only indirectly aligned with the metrics that matter in downstream quantitative tasks. FiRE addresses the first challenge by encoding each candlestick bar as a 256-bit binary token, defining an implicit code space, and modeling next-token generation with conditional flow matching. This replaces vocabulary-sized categorical prediction with joint generation in the binary token space, enabling representation capacity to scale efficiently. For task-specific alignment, we introduce ReFLO (Rejected-Future Flow Optimization), which uses verifiable financial rewards to select low-scoring futures and constructs repulsive token-level regression targets, without differentiating through rewards or evaluating policy likelihoods. Extensive experiments demonstrate FiRE's superior performance across a range of quantitative finance tasks. FiRE improves return forecasting IC by 44% over the strongest baseline in the zero-shot setting, with the gain increasing to 51% after ReFLO post-training. It also achieves a 33% higher discriminative score than baselines for synthetic candlestick generation and delivers higher risk-adjusted returns in investment simulations.

Primary area:
learning on time series and dynamical systems
Keywords:
financial foundation models, time series foundation models, generative pretraining, flow matching, preference optimization

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_rxNCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces FiRE, a financial time-series foundation model that combines generative pretraining with task-specific reward-guided adaptation. It addresses the cost of scaling discrete market representations and the mismatch between generative objectives and downstream financial metrics. FiRE encodes OHLCVA candlesticks into binary tokens using lookup-free quantization and combines a causal Transformer with conditional flow matching to generate all bits of each future token jointly. The authors further propose ReFLO, which scores sampled futures using verifiable financial rewards and constructs repulsive regression targets from continuations with below-average rewards, without requiring reward gradients or policy likelihoods. The framework is evaluated on financial forecasting across markets and frequencies, synthetic candlestick generation, and investment backtesting, with ablations examining representation capacity, generation mechanisms, and post-training feedback.

Critical strengths
  1. The architecture provides a practical approach to increasing market representation capacity. Combining binary tokens with joint flow-based generation avoids a vocabulary-sized classifier and sequential prediction within each token, while retaining autoregressive modeling across market timesteps. The tokenizer's candlestick-validity and directional objectives also incorporate relevant structure from financial data.

  2. ReFLO offers a flexible way to incorporate financial task feedback into a flow-based generator. Its repulsive regression updates do not require differentiable rewards or explicit policy likelihoods, making the approach compatible with the proposed architecture. Different reward functions allow the same pretrained model to be adapted separately to price, return, and volatility prediction.

  3. The empirical evaluation covers several aspects of financial modeling, including forecasting, synthetic data generation, and investment simulation with transaction costs. Comparisons include task-specific models, general time-series foundation models, and financial foundation models. The ablations on token width, prediction heads, sampling, and feedback signals provide useful evidence about the design choices and the gap between reconstruction quality and downstream forecasting performance.

Critical weaknesses
  1. The connection between ReFLO's local surrogate and the financial objective remains insufficiently established. Equations (10)–(11) show how the regression target encourages a denoising prediction to move away from a rejected token, but do not establish how this changes the probability mass assigned to low-reward continuations under the full sampler. Since the update acts on continuous token predictions while rewards are evaluated on complete decoded trajectories, the relationship between these two levels is important to the method's justification. An analysis or controlled experiment connecting the local update to changes in the generated distribution and expected reward would strengthen this contribution.

  2. The data splits and model-selection procedures are not specified clearly enough to assess out-of-sample validity. Appendix A.1.3 provides chronological splits for pretraining and standard forecasting, but refers to a separate fixed split for the main benchmark whose evaluation dates are not given in Appendix H.1. The date ranges used for forecasting and investment post-training, and the corresponding checkpoint-selection rules, also remain unclear. These details are particularly important because ReFLO computes rewards from realized futures. In addition, the main forecasting and investment tables report point estimates without variability across independent runs or confidence intervals accounting for temporal dependence, making the stability of the gains difficult to assess.

  3. The comparisons leave some ambiguity about the sources of the reported improvements. The post-training experiments compare different feedback signals, but lack conventional task fine-tuning or alternative reward-based adaptation from the same initialization with matched data and comparable compute. This limits the ability to distinguish the contribution of ReFLO's update rule from that of additional task supervision. In the investment experiment, FiRE and the baselines also construct ranking signals from different forecast horizons. Although the authors explain this choice and keep the trading protocol fixed, results at common horizons and an explicit validation-based horizon-selection procedure would help separate model improvements from signal-design choices.

Questions for authors
  1. Averaging continuous endpoints and then thresholding each bit need not preserve the learned joint token distribution. Have the authors examined how this aggregation affects dependencies between bits and the validity of decoded candlesticks as the sampling budget changes? Such an analysis would help explain the interaction between aggregation, the learned representation, and forecasting performance.

  2. In Table 2, FiRE achieves a better discriminative score than Kronos, while its train-on-synthetic, test-on-real forecasting scores are slightly lower. What explains this difference? In particular, which financial properties, such as heavy-tailed returns and volatility clustering, are preserved by the generated data, and which temporal dependencies appear most relevant to its predictive usefulness?

  3. Since price, return, and volatility adaptation use separate rewards, do these objectives exhibit positive transfer or interfere with one another? Evaluating each adapted checkpoint across the other forecasting tasks and synthetic-data generation would help clarify whether task-specific gains preserve the pretrained model's broader financial modeling capabilities.

8 comments
AuthorsOpenReview ↗

Response to Reviewer rxNC (1/6): ReFLO and Trajectory Rewards

Part 1/6 — W1: Local updates and trajectory rewards

Thank you for your careful review and constructive suggestions. We address your three weaknesses and three questions in six numbered replies, with targeted experiments and explanations.

Both new experiments use price-path forecasting RL tasks. Each complete decoded OHLC trajectory is scored by its temporal Pearson and Spearman correlations with the realized future, averaged equally over the two correlations and valid OHLC fields. Expected reward is estimated by scoring individual trajectories and then averaging their rewards.

1. From the local surrogate to trajectory reward. Following Appendix F.3, we connect ReFLO's local regression to reward-guided reweighting of complete trajectories. Fix a history hh and its realized future. Let p0(S∣h)p_0(S\mid h) be the current distribution of complete continuations, Rh(S)R_h(S) their decoded trajectory reward, and ah(r)=E[ω−∣Rh(S)=r]a_h(r)=\mathbb{E}[\omega^-\mid R_h(S)=r] the expected negative advantage. Since ah(r)a_h(r) is nondecreasing in reward, the reference distribution

=\frac{p_0(S\mid h)\exp\\!\left(\eta a_h(R_h(S))\right)}{Z_\eta(h)}$$ gives lower-reward continuations a stronger penalty and higher-reward continuations more relative weight. Here $\eta>0$ controls the strength of reweighting, and $Z_\eta(h)$ normalizes the distribution. This reweighting improves the full-trajectory reward distribution. Its expected reward is nondecreasing, while the probability below any fixed reward threshold $r_h$ is nonincreasing: $$\frac{d}{d\eta}\mathbb{E}_{q_\eta}[R_h] =\operatorname{Cov}_{q_\eta}\\!\left(R_h,a_h(R_h)\right)\geq0, \qquad \frac{d}{d\eta}\Pr_{q_\eta}(R_h\leq r_h)\leq0.$$ The connection to ReFLO is through the corresponding conditional denoiser. Let $u$ collect the history, generated prefix, noisy token and flow time, and let $e$ be the clean token. In the population analysis of Appendix F.3, with $f_0(u)=\mathbb{E}_{p_0}[e\mid u]$ and $a=a_h(R_h(S))$, the ReFLO regression solution and the denoiser induced by $q_\eta$ satisfy $$f_{\mathrm{ReFLO},\eta}(u) =f_0(u)+\eta\operatorname{Cov}_{p_0}(e,a\mid u) =f_{q_\eta}(u)+O(\eta^2).$$ **ReFLO therefore matches both the direction and magnitude of the leading denoiser change induced by reward-improving trajectory reweighting.** This connects the local token update to a change in the complete-trajectory distribution, whose expected reward and low-reward probability follow the directions above. The complete theoretical derivation is provided in Appendix F.3. **2. One actual update.** Across three runs, we evaluated zero, half, and full strength of the first accepted parameter update on 124 feedback histories per run, using 512 fresh trajectories per history and strength. Averaged across the three runs, the full step increased mean reward by **0.001052** [0.000074, 0.002031] and reduced probability at or below each history’s independently calibrated initial 20th-percentile reward by **0.199 percentage points** [0.052, 0.346]. Half-step point estimates lay between the endpoints. All comparisons rerun Gaussian-initialized Euler16 with $K = 1$, token thresholding, each continuation’s own autoregressive prefix, and the frozen decoder, using paired source noise. **3. Distribution after 1 and 12 updates.** Separately, we evaluated all three runs at updates 0, 1, and 12 on a fixed benchmark of 768 histories across eight market–frequency settings. Each history’s quintile boundaries were fixed using 128 independent baseline trajectories. Another 128 trajectories per model/history measured probabilities in these unchanged ranges. | Fixed baseline reward range | Before (%) | After 1 update (%) | After 12 updates (%) | Change (12−0), pp [95% MC interval] | |---|---:|---:|---:|---:| | 0–20% | 20.41 | 20.27 | 20.28 | −0.14 [−0.42, +0.14] | | 20–40% | 19.60 | 19.61 | 18.49 | −1.11 [−1.41, −0.81] | | 40–60% | 19.71 | 19.69 | 17.71 | −2.00 [−2.31, −1.70] | | 60–80% | 19.64 | 19.86 | 18.82 | −0.83 [−1.13, −0.53] | | 80–100% | 20.64 | 20.58 | 24.71 | +4.08 [+3.79, +4.36] | Mean trajectory rewards at updates 0, 1, and 12 were **0.037745, 0.038528, and 0.065886**. The first-step benchmark change was small and its interval includes zero: +0.000784 [−0.000300, +0.001867]. By update 12, mean reward increased in all three runs, with a mean increase of **+0.028141** [0.025610, 0.030672]. The clearest accumulated effect is increased high-reward probability and expected reward. Results give equal weight to settings and runs. All intervals are pointwise 95% paired Monte Carlo intervals. The trajectory-reweighting derivation and its assumptions are now presented in **Appendix F.3 (Connection to Trajectory Rewards)**.
AuthorsOpenReview ↗

Response to Reviewer rxNC (2/6): Data Separation and Evaluation Reliability

Part 2/6 — W2: Temporal separation and evaluation reliability

W2: Temporal separation, model selection, and evaluation reliability.

Thank you for highlighting the need to make these procedures explicit. We apologize for the unclear description of data splits and model selection in the original manuscript and have clarified these details in the paper.

Data separation and model selection. Both forecasting and investment ReFLO use the same historical training corpus as pretraining, with a cutoff of February 29, 2024. FiRE uses March–June 2024 for validation and selects checkpoints using the corresponding validation metrics. The main forecasting benchmark and investment evaluation both cover July 2024–December 2025, totaling 18 months. Importantly, computing rewards from “realized futures” does not mean using future information from the test period: these are observed continuations of historical training windows, entirely contained within the training split, with all reward targets ending before the training cutoff. No forecasting or investment test outcomes are used for parameter updates or checkpoint selection. We have clarified these chronological splits in Appendices A.1.3, H.1, and I.1, explicitly adding the evaluation dates of the fixed 62-task benchmark in H.1.

Forecasting reliability. We assess FiRE and task-specific FiRE + ReFLO on the central price-forecasting (PCC/SCC) and return-forecasting (IC/RankIC) tasks through 32 paired comparisons against competitive baselines. We use a paired moving-block bootstrap over evaluation dates, jointly resampling the same date blocks across models and market–frequency settings to account for temporal and cross-sectional dependencies. Within each bootstrap sample, we recompute the forecasting metrics and average equally across the 62 settings. We report pointwise 95% confidence intervals for score differences and apply Bonferroni correction across all 32 comparisons. This procedure is now specified in Appendix H.5.

The table shows comparisons against the strongest evaluated baseline for each metric. Entries are improvements over the baseline with 95% confidence intervals; positive values indicate better performance.

MetricBaselineFiRE improvement [95% CI]FiRE + ReFLO improvement [95% CI]
PCCKronos Base0.00230 [0.00158, 0.00302]0.00700 [0.00477, 0.00923]
SCCKronos Base0.00080 [0.00050, 0.00110]0.00740 [0.00457, 0.01023]
ICMoirai 2.0 Small0.03330 [0.02013, 0.04647]0.03820 [0.02257, 0.05383]
RankICKronos Base0.01640 [0.01366, 0.01914]0.02010 [0.01670, 0.02350]

All 32 improvements are positive, all 95% confidence intervals exclude zero, and the Bonferroni-adjusted pp-values are at most 3.2×10−43.2\times10^{-4}. These results provide clear statistical support for the gains over the strongest evaluated baselines. Additional evaluations across different test periods also retain the performance advantages, supporting their temporal stability. The complete paired comparisons appear in Appendix H.5 and Figure H.1.

We additionally conducted three independent post-training runs for each of the price, return, and volatility objectives; all three runs per objective improve their targeted metric. Full training curves, cross-run means, and standard deviations appear in Appendix F.6, Figure F.1, and Table F.2.

Investment reliability. For each model, we select the forecast horizon with the highest validation ESR among 5, 15, and 30 trading days (15 days for FinCast and Kronos Base; 30 days for Sundial, FiRE, and FiRE + ReFLO), and fix it for test evaluation. At the selected horizon, we evaluate Sundial, Kronos Base, FiRE, and FiRE + ReFLO using five sampling seeds, keeping the checkpoint, sampling budget, test period, and trading protocol fixed. We report means and sample standard deviations to quantify sampling-induced variability. FinCast’s inference is deterministic, so its results are reported without a standard deviation. Higher AER and ESR indicate better performance.

ModelAER (%): mean ± SDESR: mean ± SD
FinCast9.381.00
Sundial10.25 ± 1.030.83 ± 0.10
Kronos Base11.03 ± 2.040.98 ± 0.17
FiRE13.65 ± 1.101.20 ± 0.09
FiRE + ReFLO16.98 ± 1.401.50 ± 0.12

FiRE outperforms all three external baselines in mean AER and ESR, with each baseline evaluated at its validation-selected horizon. ReFLO further improves AER by 3.33 percentage points and ESR by 0.30 over pretrained FiRE, corresponding to relative gains of 24.4% and 25.0%. These repeated evaluations confirm FiRE’s advantage in average investment performance and the additional gains from ReFLO.

AuthorsOpenReview ↗

Response to Reviewer rxNC (3/6): ReFLO Attribution and Investment Horizons

Part 3/6 — W3: ReFLO attribution and investment horizons

W3: Attribution of the gains.

1. ReFLO's contribution beyond additional training and task supervision

Thank you for this suggestion. Using return forecasting as a representative task, ReFLO outperforms continued pretraining, both supervised adaptation baselines, and all three alternative reward-based methods. All methods start from the same pretrained FiRE checkpoint and use the same training windows with matched total window exposure. Configurations and checkpoints are selected by validation performance and evaluated under the same test settings. Reward-based methods additionally use the same return reward and candidate-sampling budget.

Continued pretraining retains the original flow-matching objective, testing whether additional generative training alone produces the gains. For direct task supervision, we train the same small MLP return head either with the backbone frozen or with both the backbone and head updated. This tests whether changing the readout or adapting the representation through return regression is sufficient. We also compare AWM [1], Flow-GRPO [2], and DiffusionNFT [3] with ReFLO using the same return reward and evaluation protocol.

ReturnScore is the average of return IC and RankIC; higher is better.

MethodReturnScore ↑
Pretrained FiRE0.0765
Continued pretraining0.0710
Frozen backbone + return head0.0796
Task-supervised fine-tuning0.0788
AWM0.0803
Flow-GRPO0.0670
DiffusionNFT0.0812
ReFLO0.0840

Both supervised controls improve over pretrained FiRE, with head-only training slightly outperforming joint backbone–head fine-tuning. AWM and DiffusionNFT also improve the score, while continued pretraining and Flow-GRPO reduce it. ReFLO exceeds all of these alternatives, showing that its gains go beyond those obtained from additional generative training, direct return supervision, or the alternative updates using the same reward.

ReFLO performs best: its ReturnScore gain over pretrained FiRE is 9.80%, compared with 6.14% for DiffusionNFT, the strongest alternative—a 59.6% larger improvement. Its advantage over the supervised controls and methods using the same reward supports the contribution of ReFLO's update rule beyond additional training and task supervision.

2. Model selection and forecast horizons in the investment experiment

Thank you for raising this concern. For supervised baselines that require training, we already use validation-based model selection: training covers January 2011–February 2024, and validation covers March–May 2024. For each architecture, we select the configuration with the lowest validation return MSE and apply early stopping. This procedure is explicitly described in Appendix I.1.

To directly test whether the investment advantage comes from different forecast horizons, we added comparisons at common horizons of 5, 15, and 30 trading days for all models, keeping the trading protocol fixed. Both FiRE and FiRE + ReFLO outperform the evaluated external baselines in AER and ESR at all three common horizons.

The table below shows the strongest external baseline at each horizon alongside the two FiRE models. Each entry reports AER (%) / ESR; higher is better for both metrics.

Forecast horizonStrongest external baselineFiREFiRE + ReFLO
5 trading daysFinCast: 8.01 / 0.868.895 / 0.90310.588 / 1.022
15 trading daysKronos Base: 11.16 / 1.08012.740 / 1.19912.701 / 1.420
30 trading daysSundial: 10.38 / 0.83013.77 / 1.20317.50 / 1.50

FiRE's advantage persists even at the five-day horizon originally used by the baselines. FiRE + ReFLO achieves the best ESR at all three horizons, further showing that its risk-adjusted return advantage is not confined to the original 30-day setting. The original difference in forecast horizons therefore does not explain this advantage.

The complete model-by-horizon comparison is now included in Appendix I.5 and Table I.2. The common trading protocol and supervised model-selection procedure are specified in I.1 and I.3.

References

[1] Advantage Weighted Matching: Aligning RL with Pretraining in Diffusion Models. ICML 2026.

[2] Flow-GRPO: Training Flow Matching Models via Online RL. NeurIPS 2025.

[3] DiffusionNFT: Online Diffusion Reinforcement with Forward Process. ICLR 2026, Oral.

Reviewer_wA7KOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a foundation model FiRE for financial market forecasting and generation, in the candlestick form. FiRE introduces two key designs, a flow-based generative pretraining with a new market tokenization and a reward-guided alignment. The proposed tokenization, based on a lookup-free quantization, enables efficiently scalable representation of candlestick bars within a relatively large but not too expensive code space. Paired with the conditional flow matching, it generates all bits of the next tokens jointly, thererby, avoiding sequential subtoken prediction. The post-training alignment ReFLO iscores sampled future trajectories with verifiable financial rewards and uses negative group-relative advantages to construct repulsive token-level targets, adapting the generator to downstream tasks without reward gradients or policy likelihoods. Experimental results show that FiRE outperforms both general-purposed and task-specific time series forecasting baselines across multiple quantitative finance tasks and generates high-fidelity candlestick sequences.

Critical strengths
  1. This paper studies a critical research problem in finance. The motivations are also clear and solid.

  2. The two key designs effectively address the challenges of appropriate tokenization and efficient task alignment. The empirical validation of token size also provides interesting insights.

  3. Extensive experimental results are encouraging, showing that FiRE performs well across a wide range of financial tasks. Economic evaluation metrics are employed beyond conventional forecasting errors and correlations, providing evidence of actual values in financial economics.

  4. Sufficient ablation studies demonstrate the effectiveness of each designed component.

Critical weaknesses
  1. I understand that reporting the optimal performance of each model is a standard practice in time series forecasting. However, it is insufficient to demonstrate whether the outperformance is significant. Significance tests, confidence intervals, and error bars should be provided.

  2. The testing period is fixed and relatively short (roughly half a year only). Financial markets shifts with distinct temporal regimes. Prediction models may perform very differently in distinct periods. Given the test set in this paper, it’s unclear whether the model successfully captures dynamic signals and perform consistently across time.

Minor:

  • Legend at the bottom in Figure 1 is a little confusing. Improve the readability by adjusting the legend position, for example, placing it vertically on the right-hand side
Questions for authors
  1. Does the outperformance remain in a different testing period?

  2. This paper provides the cutoff for the training set only and does not clearly specify the range. How many years do the training data cover?

  3. The candle validity is important. However, the validity is reported only for FiRE, but not for other baselines. Since OHLC legality is a basic property of candlestick data, please report the validity rates of forecasts and generated samples from representative baselines as well.

  4. Does the model capture accurate dependencies among OHLCVA dimensions? Could you also measure the dependencies?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wA7K (1/2)

Part 1/2 — W1–W2/Q1–Q2: Reliability and data coverage

Thank you for your thoughtful review and encouraging assessment. Your suggestions have helped us clarify the data coverage and strengthen our statistical and generation analyses.

W1: Statistical reliability.

The added tests support the significance of the forecasting gains, including the smaller improvements in price prediction. We compare FiRE and task-specific FiRE + ReFLO with four competitive baselines per forecasting task on matched evaluation windows, yielding 32 paired comparisons with 95% confidence intervals and Bonferroni-adjusted p-values.

We focus on price and return forecasting to test whether even the relatively modest gains over strong baselines are statistically supported.

Below, each entry is the score improvement over the highest-scoring evaluated baseline for that metric, followed by its 95% CI. Positive values favor FiRE.

MetricBaselineFiRE improvement [95% CI]FiRE + ReFLO improvement [95% CI]
PCCKronos Base0.00230 [0.00158, 0.00302]0.00700 [0.00477, 0.00923]
SCCKronos Base0.00080 [0.00050, 0.00110]0.00740 [0.00457, 0.01023]
ICMoirai 2.0 Small0.03330 [0.02013, 0.04647]0.03820 [0.02257, 0.05383]
RankICKronos Base0.01640 [0.01366, 0.01914]0.02010 [0.01670, 0.02350]

All 32 improvements are positive, all intervals exclude zero, and all adjusted p-values are at most 3.2×10−43.2\times10^{-4}. Appendix H.5 and Figure H.1 report the full comparisons and uncertainty bars. We also added three training runs per post-training objective, with longer curves and cross-run standard deviations (Appendix F.6, Figure F.1, Table F.2).

W2/Q1: Evaluation period.

The evaluation covers July 2024–December 2025: 18 months, rather than six. The original December 2024 end date was a typo, which we have corrected in Appendix A.1.3; Appendix H.1 now specifies the frozen 62-task benchmark dates. The investment backtest uses the same period (Appendix I.1), with full-period investment paths in Appendix I.4 and Figure I.1.

Q2: Training-data coverage.

The pretraining corpus spans December 19, 1990–February 29, 2024, approximately 33 years overall. Coverage varies by market. Appendix A.1.1 and Table A.1 now state the overall range, and Table A.2 lists source-group start dates. The data universe and cleaning follow Kronos; FiRE's training/validation/test boundaries are specified in Appendix A.1.3.

AuthorsOpenReview ↗

Response to Reviewer wA7K (2/2)

Part 2/2 — Q3–Q4: Generation quality and presentation

Q3: Candlestick validity of baselines.

We added validity rates for representative baselines (Appendix J.1, Table J.1). A valid candle has finite OHLC values satisfying H≥max⁡(O,C)H\geq\max(O,C) and L≤min⁡(O,C)L\leq\min(O,C). Results use raw outputs without additional OHLC repair, averaged equally across market–frequency settings.

ModelCandle validity (%)
DUET79.47
Moirai 2.0 Small82.43
TimeVAE99.97
TimeGAN89.95
Diffusion-TS89.57
Kronos Base89.00
FiRE97.21

FiRE exceeds Kronos by 8.21 percentage points on forecast outputs. TimeVAE has higher validity, but its discriminative score of 0.0407 versus FiRE's 0.3933 (closer to 0.5 is better; Table 2) shows that valid candles alone do not ensure realistic sequences.

Q4: Dependencies among OHLCVA dimensions.

Thank you for this suggestion. By OHLCVA dependencies, we mean the statistical relationships among open, high, low and close prices, trading volume (V), and traded amount (A). For example, bars with larger price changes may also have higher trading volume. We test whether generated data reproduce the strength and direction of these relationships.

We use Spearman correlations, calculated across bars within each sequence. Before computing them, we express each OHLC price as Pt/Ct−1−1P_t/C_{t-1}-1, using that sequence's own preceding close, to reduce the influence of shared price levels. We examine two aspects:

  • Relationships among all six fields: all 15 combinations of two fields measured at the same time.
  • The size of price movements and trading activity: absolute close-to-close returns and relative high–low ranges, each paired with V and A, giving four relationships. This specifically tests whether larger movements accompany greater activity, since both large upward and downward moves can be associated with high volume. Relative range is (Ht−Lt)/Ct−1(H_t-L_t)/C_{t-1}.

For each relationship, market–frequency setting, and sampling seed, we average the correlations separately over real and generated sequences. Its correlation error is the absolute difference between these two averages. The table averages these errors over the relevant relationships, settings, and sampling seeds. Lower is better; zero means the measured correlations match the real-data reference.

We compare frozen FiRE and Kronos Base on stock data from the Shanghai Stock Exchange and the Taiwan Stock Exchange at 15-minute and daily frequencies, using the same 6,000 historical windows per setting and three sampling seeds.

Relationships evaluatedKronos Base: correlation error ↓FiRE: correlation error ↓
All pairs among the six fields at the same time0.07110.0419
Absolute return and relative price range versus volume and amount0.08640.0525

FiRE reduces these errors by 41.1% and 39.3%, respectively, with lower aggregate errors in every sampling run. As a concrete example, the mean absolute-return–volume correlation is 0.3755 in real data, 0.3406 for FiRE, and 0.2598 for Kronos. Range–activity relationships remain underestimated by both models on average. These results show that FiRE reproduces the measured relationships among the six fields more closely on average.

The metric definitions, evaluation protocol, and per-setting results are now included in Appendix J.3 and Table J.2.

Minor: Figure 1. We revised the legend layout to group labels with the corresponding panels and make the series easier to identify.

Thank you again for your encouraging assessment and constructive suggestions. We hope these clarifications and additional results address your concerns and strengthen your confidence in our work. We would be happy to clarify any remaining questions.

Thanks for your response and additional experiments. Overall, my concerns have been fully addressed, and I've raised my rating.

We are especially grateful for your support from the outset and your careful consideration of our responses and additional experiments. Your constructive feedback and encouraging reassessment mean a great deal to us.

Reviewer_jGx4OpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FiRE, a financial foundation model that encodes candlestick (OHLCVA) bars into 256-bit binary tokens via lookup-free quantization (LFQ) and models autoregressive next-token generation using a conditional flow-matching head. To align the pretrained generator with downstream metrics without requiring policy likelihoods or reward gradients, the authors propose ReFLO (Rejected-Future Flow Optimization), which leverages negative group-relative advantages from verifiable financial rewards to construct repulsive token-level regression targets. The framework is evaluated across 62 multi-frequency forecasting tasks, CSI1000 portfolio backtesting in Qlib, and synthetic candlestick generation. However, important baselines like XGBoost are missing.

Critical strengths
  • The work is generally well-presented, with clear figures.
  • the idea of creating detached, repulsive targets using forward-process token probes provides a practical method for adapting continuous flow heads using non-differentiable reward functions.
  • extensive experiments and the scaling ladder are used for evaluation
Critical weaknesses
  • the path-based volatility evaluation seems to penalize smooth point/median forecasts. since the volatility is not predicted directly, most baselines (eg. dlinear) use deterministic point, mean or median readouts as shown in Table H.4, their predicted trajectories are smooth and underestimate high-frequency intrabar ranges.
  • missing critical baselines, for example, LightGBM and XGBoost
  • the post-training is only shown for 36 steps in figure 1b, it would be helpful to show longer training and multi-seed variance results.
Questions for authors

please see weakness. i will be happy to raise my rating based on how my concerns are addressed

3 comments
AuthorsOpenReview ↗

Response to Reviewer jGx4

Thank you for your careful review and constructive suggestions, which have helped us strengthen the evaluation and improve the paper. We have addressed all three concerns with the following additional evidence.

W1: Fairness of volatility evaluation.

We agree that smooth point or median OHLC forecasts can suppress variability, so variance derived from these paths does not fully reflect a model’s ability to predict variance directly. To address this concern, we expanded the comparison to include dedicated variance predictors—HAR, GARCH, LightGBM, XGBoost, and DLinear-DirectVol—alongside the historical-variance baseline Historical-YZ.

All methods use the same 13,824 windows across 54 equity tasks, with matched histories, horizons, future Yang–Zhang variance targets, and QLIKE calculation. Direct variance prediction avoids the smooth-path issue raised in your review. Fitted baselines use training data for fitting and validation data for selection. FiRE and Kronos use pretrained checkpoints without volatility-specific fine-tuning; each averages the variances calculated separately from ten sampled paths.

Lower QLIKE indicates better performance.

MethodQLIKE ↓ [95% interval]
Historical-YZ3.2155 [1.4798, 6.4752]
DLinear-Path1.2158 [1.1380, 1.3156]
DLinear-DirectVol3.7639 [1.8702, 6.4920]
HAR2.1857 [1.9152, 2.4618]
GARCH1.0496 [0.9542, 1.1608]
LightGBM0.8219 [0.4701, 1.3326]
XGBoost0.6980 [0.6315, 0.7741]
Kronos1.2902 [0.9550, 1.6495]
FiRE0.3704 [0.3248, 0.4157]

Task-macro averages; 95% intervals use 20,000 paired market-cluster bootstrap resamples.

FiRE performs best: its QLIKE is 46.9% lower than XGBoost, the strongest dedicated predictor, and 71.3% lower than Kronos. The comparison with XGBoost addresses the smooth-path concern; the comparison with Kronos evaluates two pretrained generators using the same readout. Together, they show that FiRE's advantage persists against direct variance prediction and that its generated paths retain useful volatility information. The revised Appendix H.6/Table H.6 reports this comparison and clarifies the evaluation scope; A.2, Eq. (A.10) specifies the variance-averaging readout.

W2: Missing tree-based baselines.

We added LightGBM, XGBoost, CatBoost, and Random Forest for price and return forecasting. All four use the same normalized historical OHLCVA windows as FiRE, flattened into feature vectors similar to Qlib’s Alpha360 representation [1]. We train separate regressors for price and return forecasting, with one regressor per forecast step. For each baseline and forecasting objective, we evaluate candidate hyperparameter combinations on the validation set and select the configuration with the best aggregate performance across forecast steps. All steps share this configuration, which remains fixed throughout testing. PCC/SCC measure price prediction and IC/RankIC measure return prediction; higher is better.

ModelPCC ↑SCC ↑IC ↑RankIC ↑
LightGBM0.01220.01090.04890.0439
XGBoost0.00900.00840.03610.0363
CatBoost0.00270.00210.03770.0383
Random Forest−0.0053−0.00570.03240.0339
FiRE0.04910.04500.10850.0853
FiRE + ReFLO0.05380.05160.11340.0890

FiRE outperforms all four added baselines on both tasks before post-training, and ReFLO improves all four metrics further.

W3: Longer training and multiple seeds.

We use return forecasting as a representative objective and extend post-training to 72 updates with three training seeds, all starting from the same pretrained checkpoint. We evaluate every checkpoint on the same 62 tasks. ReturnScore averages IC and RankIC; higher is better. The shared pretrained score is 0.076546.

Training seed / step122436486072
Seed 10.0797130.0822800.0830550.0836500.0837460.084554
Seed 20.0797300.0800040.0806070.0823170.0839050.084251
Seed 30.0792100.0810760.0814560.0824680.0824140.083103
Mean0.0795510.0811200.0817060.0828120.0833550.083970

At step 72, mean ReturnScore is 0.083970 ± 0.000765 (sample SD), 9.70% above pretraining and 2.77% above step 36. All three seeds improve over their 36-step checkpoints, confirming gains across runs.

Appendix F.6, Figure F.1, and Table F.2 report the full curves and multi-seed results, including additional objectives.

Thank you again for your suggestions. We hope these comparisons and multi-seed results resolve your concerns, and would appreciate your reconsideration of the rating. We are happy to clarify any remaining questions.

[1] Qlib. Alpha360 feature construction, official implementation.

thanks for your time and effort on the response and additional experiments. my concerns are adequately addressed so i will raise my rating.

Thank you for your careful reassessment; we are glad the additional experiments addressed your concerns and appreciate your constructive suggestions for strengthening the evaluation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the AC

I. Acknowledgments

We sincerely thank all three reviewers for their careful reviews and constructive suggestions. We especially appreciate the time they devoted to examining our responses and additional experiments, and their willingness to reassess our work in light of the new evidence. Their thoughtful engagement throughout the discussion has helped us strengthen the paper. We also thank the Area Chair for overseeing the review and discussion.

II. Strengths Highlighted by the Reviewers

  • Scalable representation and generation. Binary market tokens and joint flow-based generation expand representation capacity without a vocabulary-sized classifier or sequential within-token prediction (wA7K, rxNC).
  • Practical task adaptation. ReFLO incorporates financial task feedback into the generator without reward gradients or policy likelihoods, providing a flexible, practical way to adapt the pretrained model to different financial objectives (jGx4, rxNC).
  • Broad evaluation and financial value. Forecasting, generation, and investment experiments cover diverse financial tasks, supported by scaling and component ablations (all three reviewers). Economic evaluation further demonstrates practical value beyond conventional forecasting errors and correlations (wA7K).
  • Clear presentation and solid motivation. The paper is clear and easy to follow (all three reviewers), and studies an important financial problem with clear, solid motivation (wA7K).

III. Additional Evidence from the Discussion

Our additional experiments show that FiRE outperforms the added tree-based forecasting models and dedicated variance predictors. It also maintains its investment advantage at common forecast horizons (jGx4, rxNC). Statistical tests and multi-seed evaluations further support the reliability of the reported gains. We have also clarified the chronological data splits and validation-based selection procedures (jGx4, wA7K, rxNC).

The discussion also deepened our understanding of the framework. Controlled comparisons show that ReFLO outperforms continued pretraining, supervised adaptation, and alternative reward-based methods. Theoretical analysis and trajectory-level experiments connect its local updates to improved financial rewards (rxNC). We also examined aggregation and generated-data quality (wA7K, rxNC). Cross-task evaluation revealed an interesting pattern: price and return adaptation benefit each other but impair volatility forecasting. Volatility adaptation, in contrast, improves all three forecasting capabilities and the average predictive utility of synthetic data (rxNC).

IV. Manuscript Revisions

The revised manuscript clarifies data splits and model-selection procedures (Appendices A.1.3, H, I), adds statistical reliability and multi-seed training analyses (H.5, F.6), dedicated volatility and common-horizon investment comparisons (H.6, I.5), and analyses of ReFLO's theory, aggregation, generation quality, and cross-task transfer (F.3, C.7, J, F.7).

We again thank the reviewers and Area Chair for their time and thoughtful consideration, and remain happy to clarify any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.