STRIDE Moves Market Sentiment

Sujay Rittikar, Sheela Ramanna
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:1151-1156, 2026.

Abstract

Aspect-based sentiment analysis in the financial domain requires models to reason over sparse, entity-centric signals while remaining robust to linguistic variability and conflicting cues. We introduce STRIDE, a reinforcement learning framework that reformulates keyword selection as a sequential decision-making problem, integrating Directional Stimulus Prompting (DSP) with stable reward-driven policy optimization. To address the instability of sparse, high-variance reward signals, STRIDE incorporates exponential moving average (EMA) smoothing into the REINFORCE objective, enabling more reliable gradient estimates for policy learning. We evaluate STRIDE on two benchmark financial sentiment datasets: SEntFiN 1.0 and FinEntity. On SEntFiN 1.0, STRIDE achieves state-of-the-art F1-score (0.946) and near state-of-the-art accuracy (0.950). On FinEntity, STRIDE exceeds the previous state-of-the-art F1-score by 4.2%, achieving state-of-the-art performance on both accuracy (0.942) and F1-score (0.933). Across both datasets, the results demonstrate that EMA-smoothed rewards provide consistent improvements of 2.6% to 4.3% F1 relative to unsmoothed baselines, validating the effectiveness of stability-aware reward formulation for financial aspect-based sentiment analysis. The source code for reproducibility is available at: https://github.com/sujayrittikar/stride_sentiment_analysis.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-rittikar26a, title = {STRIDE Moves Market Sentiment}, author = {Rittikar, Sujay and Ramanna, Sheela}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {1151--1156}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/rittikar26a/rittikar26a.pdf}, url = {https://proceedings.mlr.press/v318/rittikar26a.html}, abstract = {Aspect-based sentiment analysis in the financial domain requires models to reason over sparse, entity-centric signals while remaining robust to linguistic variability and conflicting cues. We introduce STRIDE, a reinforcement learning framework that reformulates keyword selection as a sequential decision-making problem, integrating Directional Stimulus Prompting (DSP) with stable reward-driven policy optimization. To address the instability of sparse, high-variance reward signals, STRIDE incorporates exponential moving average (EMA) smoothing into the REINFORCE objective, enabling more reliable gradient estimates for policy learning. We evaluate STRIDE on two benchmark financial sentiment datasets: SEntFiN 1.0 and FinEntity. On SEntFiN 1.0, STRIDE achieves state-of-the-art F1-score (0.946) and near state-of-the-art accuracy (0.950). On FinEntity, STRIDE exceeds the previous state-of-the-art F1-score by 4.2%, achieving state-of-the-art performance on both accuracy (0.942) and F1-score (0.933). Across both datasets, the results demonstrate that EMA-smoothed rewards provide consistent improvements of 2.6% to 4.3% F1 relative to unsmoothed baselines, validating the effectiveness of stability-aware reward formulation for financial aspect-based sentiment analysis. The source code for reproducibility is available at: https://github.com/sujayrittikar/stride_sentiment_analysis.} }
Endnote
%0 Conference Paper %T STRIDE Moves Market Sentiment %A Sujay Rittikar %A Sheela Ramanna %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-rittikar26a %I PMLR %P 1151--1156 %U https://proceedings.mlr.press/v318/rittikar26a.html %V 318 %X Aspect-based sentiment analysis in the financial domain requires models to reason over sparse, entity-centric signals while remaining robust to linguistic variability and conflicting cues. We introduce STRIDE, a reinforcement learning framework that reformulates keyword selection as a sequential decision-making problem, integrating Directional Stimulus Prompting (DSP) with stable reward-driven policy optimization. To address the instability of sparse, high-variance reward signals, STRIDE incorporates exponential moving average (EMA) smoothing into the REINFORCE objective, enabling more reliable gradient estimates for policy learning. We evaluate STRIDE on two benchmark financial sentiment datasets: SEntFiN 1.0 and FinEntity. On SEntFiN 1.0, STRIDE achieves state-of-the-art F1-score (0.946) and near state-of-the-art accuracy (0.950). On FinEntity, STRIDE exceeds the previous state-of-the-art F1-score by 4.2%, achieving state-of-the-art performance on both accuracy (0.942) and F1-score (0.933). Across both datasets, the results demonstrate that EMA-smoothed rewards provide consistent improvements of 2.6% to 4.3% F1 relative to unsmoothed baselines, validating the effectiveness of stability-aware reward formulation for financial aspect-based sentiment analysis. The source code for reproducibility is available at: https://github.com/sujayrittikar/stride_sentiment_analysis.
APA
Rittikar, S. & Ramanna, S.. (2026). STRIDE Moves Market Sentiment. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:1151-1156 Available from https://proceedings.mlr.press/v318/rittikar26a.html.

Related Material