How Self‑Supervised HINTS Extracts Human Insights from Time Series to Boost Forecast Accuracy

Background

Time‑series data pervade critical domains such as finance, economics, and transportation, where accurate forecasting is essential. These series embed human‑driven influences—decision‑making, sentiment, collective psychology—that manifest as fingerprints in the observed signals.

Recent works have tried to improve forecasts by ingesting external sources (news, social media) to capture these human factors, but doing so incurs substantial financial, computational, and practical costs.

Problem Definition

The authors hypothesize that the impact of external human factors is already reflected in the raw time‑series, particularly in the residual component after traditional decomposition. The goal is to extract these latent human factors without any external data, enabling interpretable and behavior‑aware predictions.

Method

Overall Workflow

HINTS consists of two training stages:

Stage 1 – Decomposition and Human‑Factor Extraction

Decompose the raw series into trend, seasonal, and residual Rᵢ(t) , treating the residual as the primary signal containing human dynamics.

Feed Rᵢ(t) into a neural extractor constrained by the Friedkin‑Johnsen (FJ) opinion‑dynamics model, producing a representation Hᵢ(t) (the human factor). This extractor is trained self‑supervised and frozen after Stage 1.

Stage 2 – Attention Modulation and Prediction

Apply a lightweight attention network to Hᵢ(t) , yielding attention scores Aᵢ(t) that highlight behavior‑relevant signals.

Fuse the modulated signal Aᵢ(t) with the original input Xᵢ(t) and feed the result into any backbone predictor (e.g., DLinear, PatchTST, TimeMixer).

Why the Friedkin‑Johnsen Model?

The FJ model balances two forces:

Self‑dynamics : persistence of an entity’s own state or bias.

Social influence : impact from other variables or external shocks.

Unlike simpler consensus models (e.g., DeGroot), FJ permits both self‑reinforcement and peer influence, making it a suitable inductive bias for modeling the interplay of personal memory and collective behavior in residuals.

Stage 1 Details

For each variable i and time step t , a lightweight network f_θ (often a linear layer) generates the human factor Hᵢ(t) . The update rule, inspired by FJ, enforces:

Hᵢ(t+1) = (1‑α)·Hᵢ(t) + α·(Σ_j w_{ij}·Rⱼ(t) + β·Rᵢ(t) + μ·̄Rᵢ(t))

where α controls the blend of self‑memory, social influence (weighted sum of other residuals), and a dynamic bias term (moving‑average of residuals). A self‑supervised loss aligns the learned Hᵢ(t) with these constraints.

Stage 2 Details

A convolutional attention module k_θ computes attention maps Aᵢ(t) from Hᵢ(t) . The original series is modulated: X'ᵢ(t) = Xᵢ(t) + γ·Aᵢ(t) ⊙ Xᵢ(t) where γ is a hyper‑parameter controlling modulation strength. The modulated series X'ᵢ(t) is then passed to the predictor g_θ to obtain the forecast.

Training Procedure

The framework is trained jointly in two phases. Stage 1 minimizes the FJ‑based self‑supervised loss, learning human factors independent of any downstream target. Stage 2 adds the prediction loss, allowing the human factors to be fine‑tuned for forecasting performance while preserving their interpretability thanks to the frozen FJ constraints.

Experiments

Datasets

Fourteen datasets are used, including public benchmarks (CDC flu reports, multi‑currency exchange rates, San Francisco highway occupancy, household electricity consumption, PeMS traffic speeds) and financial series (technology stocks, S&P 100/500) covering Jan 2020–Apr 2025.

Baselines

Models using only raw series: DLinear, PatchTST, TimeMixer.

External‑data model: “From News to Forecast” that incorporates filtered news via a large language model.

Main Results

Across all benchmark datasets, HINTS consistently outperforms the backbone models and their variants at multiple horizons. Notable gains include:

28.9 % improvement on PeMS traffic data by modeling collective movement patterns.

12.7 % improvement on exchange‑rate data by capturing market sentiment.

Up to 32.6 % gain on disease data where behavioral influence is weaker, demonstrating strong generalization.

On real‑world financial series, HINTS yields up to 15.2 % higher accuracy, with especially large benefits for long horizons (h = 48, 60), indicating that the extracted human factors capture latent dynamics such as sentiment, coordination, and herd behavior.

Case Study & Interpretability

Attention maps for major tech stocks (AMZN, GOOGL, NVDA) align closely with significant market events, often anticipating trends by one to two days, confirming that HINTS can reveal meaningful human‑driven signals without external inputs.

Ablation Study

On the PeMS‑08 dataset (h = 720) using TimeMixer, removing the entire FJ constraint (baseline w/o L_FJ) yields the highest error, proving the necessity of the self‑supervised component. Excluding either the social‑influence term or the combined self‑memory & dynamic‑bias term also degrades performance, highlighting their complementary roles.

Comparison with External‑Data Model

Although “From News to Forecast” benefits from rich exogenous signals, HINTS still surpasses the backbone models and narrows the gap, especially on traffic and electricity datasets where external data provide limited additional value. This demonstrates that HINTS can recover much of the advantage of external sources purely from the intrinsic series.

Sensitivity to γ

Increasing the modulation weight γ generally improves performance on exchange‑rate and PeMS‑04 datasets, confirming that stronger emphasis on the learned human factors yields more informative forecasts.