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

AI summary

Overview

• Research area: Time-series forecasting with an emphasis on human behavioral modeling; specifically, self-supervised representation learning and sociological opinion-dynamics modeling applied to machine learning.
• Technical level: Intermediate. The paper assumes familiarity with time-series decomposition (trend/seasonality/residual), attention mechanisms, and standard forecasting backbones, but the core idea is explained through an accessible sociological analogy.
• Scope: The paper proposes HINTS, a two-stage self-supervised framework that extracts a latent "Human Factor" from time-series residuals using the Friedkin-Johnsen opinion dynamics model as a structural constraint, then injects that factor into existing forecasters as an attention-based modulation — all without news, social media, or sentiment indices.

What This Paper Is About

Many recent forecasting models try to capture human psychology — market sentiment, herding, panic — by pulling in external data such as news articles or social media feeds. That is expensive to collect, clean, and compute over. HINTS argues that these same human influences are already embedded inside the time-series itself, specifically in the residual component that standard decomposition treats as random noise. The goal is to extract those latent human-driven dynamics endogenously and use them to improve forecasts.

Key Contributions

1. Reframing residuals as behavioral signal. The authors reconceptualize the residual component left over after decomposition as a carrier of human-driven dynamics, and propose a self-supervised framework to extract those patterns without any external data.
2. A novel extraction method grounded in opinion dynamics. They introduce a methodology for learning the "Human Factor" by constraining a neural extractor with the Friedkin-Johnsen opinion dynamics model, decomposing the constraint into Social Influence, Self-Memory, and Dynamic Bias terms.
3. Modular integration into existing forecasters. The learned Human Factor is converted into a soft attention map that modulates the input to a downstream forecasting model, and the authors show this works when plugged into three state-of-the-art backbones.
4. Empirical and interpretability validation. They report consistent accuracy gains across nine real-world datasets and use case studies and ablations to argue that the extracted factors align semantically with real-world events.

Main Findings

• Consistent improvement across backbones. HINTS is evaluated on top of DLinear (Zeng et al. 2023), PatchTST (Nie et al. 2022), and TimeMixer (Wang et al. 2024a), and improves forecasting accuracy across all three on the benchmark datasets and horizons tested.
• Headline gains by domain. The paper reports up to 28.9% improvement on PeMS traffic data, 12.7% on the Exchange dataset, up to 32.6% on the Illness dataset, and up to 15.2% on real-world financial datasets (Tech Stocks, S&P 100/500, Retail Stocks).
• Gains are larger at long horizons. Improvements are particularly evident at longer horizons (h = 48, 60), where compounding errors typically occur.
• Residuals are not noise. The framework's success is presented as evidence that what is traditionally dismissed as stochastic residual fluctuation encodes structured information reflecting human psychology.
• Competitive against an external-data method. Compared against From News to Forecast (Wang et al. 2024b), which uses external news and weather inputs, HINTS operates with no external inputs yet reaches accuracy levels that rival it in several settings. For example, on Exchange, HINTS (PatchTST) reports 0.44 MSE (×10⁴) and 4.91 MAE (×10³) versus News to Forecast's 0.42 and 4.83; on Electricity, HINTS (PatchTST) reports 80.18 MSE (×10⁻³) and 167.7 MAE versus 78.62 and 180.9.
• Ablations confirm both components matter. On PeMS 08 at h = 720 with TimeMixer, the full model achieves 0.243 MSE / 0.336 MAE. Removing Social Influence raises error to 0.325 / 0.375; removing both Self-Memory and Dynamic Bias yields 0.342 / 0.401; removing the entire FJ loss yields the worst result at 0.385 / 0.402.
• Attention maps align with real events. Visualizations for AMZN, NVDA, and GOOGL show the extracted Human Factor attention aligning with significant market trends, often one to two days in advance of the events, despite being trained only on time-series data.
• Larger γ tends to help. Sensitivity analysis on Exchange (h = 336) and PEMS04 (h = 24) shows that larger values of the Human Factor weighting parameter γ generally lead to improved forecasting performance.

Methodology in Plain English

HINTS runs in two stages.

Stage 1 — Extract the human factor. The raw multivariate time series is decomposed with STL into trend, seasonality, and residual. The residual is treated as the interesting part. A lightweight neural network (the paper mentions a linear layer as an example) maps each residual value into a latent embedding called the Human Factor. To stop this embedding from becoming arbitrary, the authors impose a constraint borrowed from the Friedkin-Johnsen opinion dynamics model — a sociological model in which each agent holds an intrinsic belief but is also pulled by peers' opinions. In HINTS, each variable is an agent. The expected Human Factor update at each time step is the sum of three terms:

• Social Influence: a correlation-weighted aggregate of other variables' recent residuals and latent factors.
• Self-Memory: the variable's own recent residuals and past latent state, providing temporal inertia.
• Dynamic Bias: a rolling mean of recent residuals, acting as a slow-moving local baseline.

The three terms are weighted by β, δ, and (1 − β − δ), with λ controlling susceptibility to new signals. The self-supervised loss simply measures the gap between the learned Human Factor and this FJ-implied expected update — no labels, no forecasting targets. The extractor is then frozen.

Stage 2 — Forecast with gated human modulation. The frozen extractor produces a Human Factor from new data. A small convolutional attention network (Conv1D-Tanh-Softmax) converts it into an attention map over time. The original input is then modulated element-wise: the input plus γ times (input ⊙ attention map), where γ ∈ [0, 1] controls modulation strength. The modulated series goes into an off-the-shelf forecaster (DLinear, PatchTST, or TimeMixer), placed before the final prediction head so the original architecture is not disrupted. Stage 2 is trained with a standard MSE forecasting loss.

Why This Matters

• Impact on research: The paper challenges the prevailing assumption that residuals are unstructured noise, and it opens a route to behavior-aware forecasting that avoids the cost, latency, and coverage problems of external data pipelines. It also imports opinion-dynamics theory into deep forecasting in a concrete, testable way rather than as a purely conceptual parallel.
• Real-world applications:
  • Financial forecasting and risk: capturing investor sentiment, herding, and coordination effects in equity prices and indices without paying for news feeds or sentiment APIs.
  • Public health surveillance: the Illness dataset results suggest stronger forecasting of disease trajectories, where behavioral factors like reporting delays and compliance matter.
  • Transportation and mobility planning: PeMS and Traffic results indicate that capturing collective mobility patterns improves congestion and occupancy forecasting.
  • Energy demand forecasting: the Electricity results suggest value where consumption reflects human routines and behavioral shifts.
• Industry relevance: For firms that already have proprietary internal time-series but cannot access clean external signals, an endogenous approach is directly deployable. The plug-in design means existing production forecasters can be retrofitted rather than replaced, and the attention maps give analysts something to inspect rather than a black-box score.

Future Directions

The paper does not include a dedicated future-work section, so these are open questions the work raises rather than stated next steps:

1. Choosing λ, β, and δ. The paper describes λ as a fixed coefficient and treats β and δ as weights for social influence and self-memory, but it does not report a study on how sensitive results are to these values or whether they should be learned per dataset.
2. Relaxing the frozen-stage assumption. Stage 1's extractor is trained independently and frozen for Stage 2. Whether joint or alternating training would produce better Human Factors — and whether that would damage interpretability — is not reported.
3. Evaluating on domains with weaker behavioral signatures. The paper notes that Illness shows "weaker behavioral effects" yet still gains, but broader testing outside finance, traffic, and energy is not reported.
4. Scaling the external-data comparison. The News to Forecast comparison covers Traffic, Electricity, and Exchange in Table 2; whether the near-parity result holds across all nine datasets is not reported.

Target Audience

This paper is most useful to time-series forecasting researchers and practitioners, particularly those working in financial, economic, and urban-mobility modeling. It will also interest researchers studying behavior-aware machine learning and the application of sociological models as inductive biases in neural networks, as well as engineers who need competitive forecasting accuracy without the cost and fragility of external data pipelines. Readers should be comfortable with standard forecasting benchmarks and attention mechanisms, but the sociological framing is presented for a general machine learning audience.

Authors’ 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.