HINTS: Extraction of Human Insights from Time-Series Without External Sources

TL;DR

HINTS addresses forecasting with minimal data dependency by extracting an endogenous Human Factor from time-series residuals using the Friedkin–Johnsen dynamics as a structural inductive bias. In a two-stage, self-supervised framework, Stage 1 learns $\hat{H}_i(t)$ from $R_i(t)$ under a $L_{FJ}$ constraint, and Stage 2 modulates inputs via an attention map $\mathcal{A}_D^T$ biased by $\hat{\mathcal{H}}_D^T$ with strength $\gamma$, feeding a forecasting backbone. Across nine real-world datasets, HINTS consistently improves forecasting accuracy over strong baselines, and ablations plus case studies validate the interpretability of the extracted Human Factor and its alignment with real-world events. The approach demonstrates that residuals, traditionally dismissed as noise, encode meaningful behavioral dynamics, offering a scalable alternative to external-data methods for human-centric time-series forecasting.

Abstract

Human decision-making, emotions, and collective psychology are complex factors that shape the temporal dynamics observed in financial and economic systems. Many recent time series forecasting models leverage external sources (e.g., news and social media) to capture human factors, but these approaches incur high data dependency costs in terms of financial, computational, and practical implications. In this study, we propose HINTS, a self-supervised learning framework that extracts these latent factors endogenously from time series residuals without external data. HINTS leverages the Friedkin-Johnsen (FJ) opinion dynamics model as a structural inductive bias to model evolving social influence, memory, and bias patterns. The extracted human factors are integrated into a state-of-the-art backbone model as an attention map. Experimental results using nine real-world and benchmark datasets demonstrate that HINTS consistently improves forecasting accuracy. Furthermore, multiple case studies and ablation studies validate the interpretability of HINTS, demonstrating strong semantic alignment between the extracted factors and real-world events, demonstrating the practical utility of HINTS.

Figures (4)

• Figure 1: Comparison between conventional signal-centric modeling and our proposed human-centric approach. While conventional models learn directly from combined data, HINTS explicitly extracts human-driven latent factors from raw combined data alone --- without relying on any external sources such as news. Our Human Factor Extractor, inspired by opinion dynamics, identifies hidden human influence embedded in the signal and incorporates it into the conventional forecasting model, offering a structured and interpretable path from data to decision.
• Figure 2: HINTS operates in two stages. In Stage 1, the input time series $\mathcal{X}_D^T = \{X_i(t)\}_{i,t=1}^{D,T}$ is decomposed to obtain residuals $\mathcal{R}_D^T = \{R_i(t)\}_{i,t=1}^{D,T}$, from which a Human Factor Extractor learns human-influenced signals $H_i(t)$. In Stage 2, the extractor is frozen and re-used to obtain $\mathcal{H}_D^T = \{H_i(t)\}_{i,t=1}^{D,T}$ from new data. This factor is passed to an attention block to produce a modulation signal $\mathcal{A}_D^T = \{A_i(t)\}_{i,t=1}^{D,T}$, which conditions the downstream forecasting model to generate the final prediction $\mathcal{Y}_D^T = \{Y_i(t)\}_{i,t=1}^{D,h}$.
• Figure 3: Visualization of the extracted Human Factor $\hat{\mathcal{H}_D^T}$ based attention $\mathcal{A}_D^T$ for major tech stocks.
• Figure 4: Sensitivity analysis with respect to the Human Factor weighting parameter $\gamma$ in Stage 2 of HINTS.