Reconstructed sentiment from sparse AI news reliably precedes stock-price moves by roughly three weeks across multiple pipeline designs; stable, market-relevant sentiment indices require careful aggregation and causal reconstruction rather than only better classifiers.

Main Finding

Reconstructing a causal temporal sentiment series from sparse, redundant, and uncertain news requires a dedicated three-stage pipeline (aggregation → causal gap-filling → causal smoothing). Using this approach on AI-related news (Nov 2024–Feb 2026) the authors find a robust structural regularity: the reconstructed sentiment series exhibits a persistent ~three-week lead/lag relationship with stock prices across pipeline configurations. The paper argues that reliable deployable sentiment indicators depend as much on careful reconstruction as on classifier improvements.

Key Points

• Problem framing: treat transforming article-level probabilistic sentiment outputs into a temporal signal as a causal signal-reconstruction problem rather than a pure classification task.
• Three core reconstruction stages (strictly causal at each step):
• Aggregation onto a regular time grid (weekly/monthly) with configurable weights.
• Causal gap-filling via forward carry with optional staleness decay.
• Causal smoothing (EWMA, Kalman variants, Beta–Binomial) to reduce residual noise.
• Aggregation choices and enhancements:
  • Global vs. local (per-category) aggregation.
  • Uncertainty-aware weights: normalized entropy (Went), top-two margin (Wtop), polarity-conflict (Wpol).
  • Redundancy control via embedding-based grouping; weighting families:
    • Deduplication: ϕded(n) = n^(-α) (α in [0,1]).
    • Corroboration: ϕcor(n) = 1 for n=1, log(n) for n>1.
  • Intra-bin recency weight: causal exponential-type decay within bin.
  • Composite weight wp,i = Wunc · Wdup · Wtime (multiplicative).
• Causal gap-filling: forward carry of last observed bin scaled by non-increasing staleness function g(Δ), options include constant carry-forward or finite-horizon linear decay.
• Causal smoothing: strictly using past+present only; methods include EWMA, Kalman filters (and weighted variants that use article counts tk as proxy for observation variance), Beta–Binomial conjugate smoother.
• Label-free evaluation framework (because no longitudinal ground truth):
  • Internal diagnostics: stability (e.g., total variation), smoothing-induced lag, behavior under gap-filling.
  • Counterfactual tests: impulse causality test (enforce/verify strict causal compliance), duplicate-injection test (robustness to redundancy).
  • External plausibility: consistency checks versus stock-price dynamics (secondary plausibility filter, not proof of causal effect).
• Empirical regularity: across pipeline variants and aggregation regimes, sentiment shows a three-week lead–lag pattern with stock prices—presented as a structural regularity rather than a single predictive correlation.

Data & Methods

• Data: multi-firm dataset of AI-related news titles collected by vocabulary-based scraping; evaluation period reported is November 2024 to February 2026. Each article has a timestamp, category (vocabulary), and a fixed classifier output probability vector π = (pos, neg, neu).
• Scoring: scalar sentiment sp,i = πpos − πneg ∈ [−1, 1].
• Aggregation:
  • Time binned into regular grid {D1..DT}. For bin Dk, aggregated value is typically a weighted mean of sp,i with weights wp,i.
  • Optional local (per-category then combine) vs. global aggregation.
• Uncertainty weighting:
  • Went = 1 − normalized entropy over the 3-class output.
  • Wtop = difference between top-two probabilities (margin confidence).
  • Wpol = 1 − Upol(π) where Upol penalizes polar-conflict (high mass on pos and neg but conflicted).
• Redundancy control:
  • Compute embeddings ep,i and cluster per (bank, bin) by similarity (connected components, hierarchical, etc.).
  • Assign duplication weight Wdup = ϕ(n) where n = cluster size; choose dedup (n^-α) or corroboration (log n) families depending on whether duplicates are noise or informative.
• Gap-filling:
  • Forward carry with staleness decay g(Δ) (constant or finite-horizon linear decay).
• Smoothing:
  • EWMA: S_t = α F_t + (1−α) S_{t−1}.
  • Kalman-family filters and weighted Kalman that incorporate tk (article count) as proxy for observation reliability (heteroscedasticity).
  • Beta–Binomial conjugate smoother as an alternative probabilistic smoother.
• Evaluation:
  • Internal metrics: total variation, lag induced by smoother (information-preservation lag proxies).
  • Counterfactual tests: inject impulse/noise or duplicate documents to test strict causality and robustness.
  • External check: compare lead/lag relationships with stock prices across firms. The observed ~3-week lead/lag persisted across pipeline settings.

Implications for AI Economics

• Practical signal engineering matters: For applications in market surveillance, technology diffusion tracking, or portfolio overlays, transforming per-article classifier outputs into a causal, stable time series is crucial. Improving classifier accuracy alone is insufficient.
• Robust monitoring: Embedding-based deduplication or corroboration and uncertainty weighting meaningfully change the effective signal; choice should be dataset- and task-specific (e.g., local aggregation + deduplication when thematic coverage uneven).
• Deployment constraints: Strict causality (no lookahead) must be enforced in real-time systems; gap-filling and smoothing choices introduce trade-offs between stability and responsiveness (and add predictable lag).
• Predictive structure (caveats): The observed three-week structural lead–lag vs. prices suggests informational content in reconstructed news sentiment, but external consistency is only a plausibility check—not proof of causal market impact. Economic interpretation requires further causal identification and control for confounders.
• Label-free evaluation value: In domains lacking ground-truth longitudinal sentiment labels, the proposed diagnostics and counterfactual tests offer a practical evaluation toolkit for comparing pipeline configurations and guarding against design-induced artifacts.
• Research & policy: For AI-economics studies that use textual signals to study technology adoption, competition, or asset pricing, this work highlights (i) the need to report reconstruction choices and their induced lag/stability, and (ii) that structural regularities (e.g., consistent lead/lag) can be more informative than single correlation statistics.
• Limitations to note: No ground-truth latent sentiment series is available; the external stock-price check is a secondary plausibility test; dataset specifics (vocabulary scraping, news sourcing) and parameter choices (aggregation window, clustering threshold, decay horizon) materially affect outcomes and should be validated per application.

Summary takeaway: when using news-derived sentiment in AI economics, treat temporal signal construction as a causal reconstruction problem—use uncertainty weighting, redundancy-aware aggregation, causal gap-filling, and causal smoothing, and validate pipelines with label-free diagnostics and counterfactual tests before interpreting market associations.