Does Coverage Acceleration Add Signal Beyond Velocity in Market-News Lifecycles?

Abstract#

Production market-news systems that cluster and rank stories rely on early coverage dynamics to decide which clusters will keep growing. A natural operational question is whether acceleration in unique-title arrivals — the curvature of coverage growth — carries forecasting information beyond the current level of coverage, its heat, and its recent velocity. We test this claim with a pre-registered, leakage-safe, cluster-level temporal study on 764 story clusters from a production deduplication and clustering system (81.6 days of as-of snapshot history, March–June 2026). At age six hours, we compare a regularized logistic baseline (levels, heat, title-arrival velocity, heat velocity, cadence) against an augmented model that adds a strictly backward-looking title-arrival acceleration proxy, for predicting at least two additional unique titles by age 18 hours.

On the untouched chronological holdout (N=115, 30 positives), the acceleration proxy provides no improvement: ΔAP = −0.0121 absolute average-precision points (95% moving-block-bootstrap CI [−0.0536, +0.0125]), augmented log loss and ROC AUC are numerically worse, all four secondary randomization tests are non-significant (Holm-adjusted p = 1.0), and the pre-registered acceleration-permutation negative control fails: the observed ΔAP sits below the 95th percentile of its own permutation null (+0.0101). The primary hypothesis H1 is therefore classified NOT_SUPPORTED under the frozen decision table. The level/heat/velocity baseline itself is strongly predictive of further cluster growth (AP ≈ 0.80, AUC ≈ 0.897), so the negative result is not a failure of the outcome to be learnable — it is a scoped falsification of the incremental acceleration claim. The outcome is coverage growth as recorded by the system, which includes late-ingested (backfilled) articles in 77/115 test clusters, including all 30 positives; the exclusion robustness (ΔAP = 0.0) is consistent with the null but underpowered (5 test positives). A null result of a valid execution is the deliverable; we do not reframe it.

Hypotheses#

All hypotheses and the decision table were frozen on 2026-07-22 (with two logged pre-outcome amendments, see Analysis) before any predictor–outcome association, prediction, or performance metric was inspected.

• H1 (primary) — moderate support expected, falsifiable. Among eligible clusters at age six hours, adding the preregistered backward-only acceleration proxy to the frozen baseline improves final chronological-holdout average precision for Y = 1[unique_title_count(18h) − unique_title_count(6h) ≥ 2]. Estimand: ΔAP = AP(augmented) − AP(baseline). Null: ΔAP ≤ 0. Material threshold: ΔAP ≥ 0.02 absolute AP points.
• H2 (secondary) — moderate support expected. The augmented model has lower proper-scoring-rule loss (log loss) on the final holdout. Null: Δlogloss ≤ 0 (baseline minus augmented).
• H3 (secondary) — moderate support expected. The augmented model has higher ROC AUC on the final holdout.
• H4 (secondary family) — exploratory. Any primary-model advantage is directionally preserved under alternate labels Y = 1[Δunique_titles ≥ 1] and ≥ 3.

Decision rule for H1 (frozen): SUPPORTED only if ΔAP ≥ 0.02, the paired 95% moving-block-bootstrap lower bound > 0, augmented test log loss is no worse than +0.01, and the observed ΔAP exceeds the 95th percentile of the acceleration-permutation null. Failure of the permutation control, or ΔAP ≤ 0, classifies H1 as NOT_SUPPORTED (a null result remains release-eligible if the execution package is valid). Secondaries cannot rescue H1 and are Holm-corrected across the four p-values at family-wise α = 0.05.

Data Provenance#

The data source for every analysis in this paper is a canonical immutable JSONL extract of the exact 764-cluster cohort, frozen on 2026-08-05 and stored as protocol/input-extract.jsonl (1,634,426 bytes, SHA-256 103c46c8…), whose canonical-cohort SHA-256 (f47d0dfb…) matches the frozen split counts. The analysis verifies both hashes before reading any feature or outcome value; a mismatch invalidates the experiment. The extract derives from the PerpsTrader production news database (/home/d/PerpsTrader/data/news.db, 125,662 news_articles, 11,445 story_clusters, 13.5M cluster_heat_history rows) — real production system records, not simulated. The SCOUT snapshot's physical database hash could not be reproduced later because the live append-only database had drifted, so the freeze amendment made the extract the sole experiment input. Data provenance is therefore REAL production records via frozen immutable extract (reliability of the records as database entries: RELIABLE after hash verification; as measurements of external ground truth: not error-free).

Analysis#

Design#

Unit: one story_clusters.id. Eligibility requires parseable ISO-8601 first_seen; a deterministic nearest-row snapshot within ±1 hour of each target age (3h, 6h, 18h), ordered by (absolute distance, timestamp, id); strictly increasing selected timestamps; monotone nondecreasing unique-title counts across 3h→6h→18h; finite required fields (missing heat velocity retained via train-fitted imputation plus a missingness flag). 764 of 3,621 clusters with any history qualified, spanning 2026-03-21 → 2026-06-11 (81.6 days). No cluster was removed after seeing labels, predictions, residuals, or influence. The 6h→18h horizon was fixed for snapshot density (all three ±1h windows exist for 764 clusters) and the system's ranking/escalation cadence; it is a design choice, not a tuned parameter.

Features#

All features are computable strictly from the selected 3h/6h snapshots — no 18h or later field enters fitting:

Baseline (10 inputs): log1p unique-title count at 6h; log1p article count at 6h; log1p max heat at 6h; backward title-arrival velocity V = (N6 − N3)/d36 (added by the 2026-07-27 pre-outcome amendment so the nested comparison is acceleration beyond velocity); stored heat velocity at 6h plus its missingness flag; elapsed hours a6; absolute 6h target offset; hours since the previous history row plus its missingness flag. Augmented adds exactly one feature: backward title-arrival acceleration proxy A = ((N6 − N3)/d36) − (N3/a3) — the difference between the recent 3h→6h arrival rate and the cumulative first-seen→3h average rate. Note the algebraic identity A = V − N3/a3: the augmented model is functionally the baseline with a re-weighted V coefficient plus the early cumulative rate N3/a3, so the test is of the cheap operational proxy, not a free-form curvature estimator. Prohibited fields (acceleration, predicted_heat, lifecycle_stage, peak fields, any 18h/later field, current category, article text/sentiment/prices) are enforced by the artifact's future-field tripwire.

Models, split, evaluation#

Two nested unweighted L2-regularized logistic regressions (baseline; baseline + A). Within each fit: clipping to fitting-set 1st/99th percentiles, median imputation of permitted missingness, z-standardization, unpenalized intercept, deterministic L-BFGS-B (10,000 iterations, gradient tol 1e-10, zero initialization). λ ∈ {1e-4 … 100} is chosen per model by minimum validation log loss (ties → larger λ), then the model is refit on train+validation and evaluated exactly once on the untouched test partition. Rows sorted by (first_seen, cluster_id): train 534 / validation 115 / test 115 (chronological; frozen hashes in protocol). Test class counts: 30 positive / 85 negative (26.1% prevalence; train 23.6%, validation 61.7% — a known label-shift regime handled by prevalence-standardized log loss).

# Core of the frozen artifact (research-runs/.../artifacts/analysis.py).
# Full 995-line implementation, seeds, and tripwires in the run directory.
def fit_logistic(X, y, lam):
    """Deterministic L2-regularized logistic regression via L-BFGS-B."""
    n, p = X.shape
    def neg_loglik(theta):
        w, b = theta[:p], theta[p]
        z = X @ w + b
        return (np.logaddexp(0, z).sum() - y @ z) + 0.5 * lam * (w @ w)
    theta0 = np.zeros(p + 1)
    res = minimize(neg_loglik, theta0, method="L-BFGS-B",
                   options={"maxiter": 10000, "gtol": 1e-10})
    return res.x[:p], res.x[p]

# Feature construction is as-of only: every input row uses 3h/6h snapshots.
# A = ((N6 - N3)/d36) - (N3/a3)   (augmented model adds exactly this column)
# Label Y = 1[N18 - N6 >= 2]; the future-field tripwire asserts no 18h
# feature reaches fitting. Metrics: AP, log loss, AUC on the final test
# partition; paired circular moving-block bootstrap (block 12, 10k reps,
# seed 20260722) for the Delta_AP interval.

Results#

Input integrity and determinism. Input gate: extract SHA-256 and canonical cohort SHA-256 both verified before feature construction. Determinism: two independent executions of the final artifact in a clean single-threaded process produced byte-identical results.json (SHA-256 4652ea71…, max absolute difference 0.0 vs the 1e-10 criterion). Final artifact SHA-256 3d8ccbe9…; pre-revision hashes are recorded in artifacts/provenance.json.

Primary result (test N=115, 30 pos / 85 neg):

• Paired moving-block bootstrap (block 12, 10k reps): ΔAP 95% CI [−0.0536, +0.0125] — straddles zero, leans negative; Δlogloss (base−aug) CI [−0.0166, +0.0098]; ΔAUC CI [−0.0171, +0.0096].
• Sensitivities: block 6 [−0.0517, +0.0161]; block 18 [−0.0506, +0.0111]; ordinary paired [−0.0486, +0.0145] — all straddle zero.
• Permutation negative control: FAIL. Observed ΔAP −0.0121 < null 95th percentile +0.0101 (null mean −0.0010, SD 0.0085). The frozen rule that observed ΔAP must exceed the 95th percentile is not met; the claimed incremental signal is falsified for this sample.
• H1 classification per frozen decision table: NOT_SUPPORTED (ΔAP ≥ 0.02: false; MBB lower > 0: false; proper-loss safeguard: true but irrelevant; permutation control: false).

Secondary results. All four secondary randomization p-values are far from significance; Holm-adjusted p = 1.0 for every one:

Robustness. (1) Without stored heat velocity: ΔAP = −0.01195 — unchanged conclusion. (2) Backfill exclusion (drop clusters with any >24h-backfilled article): N falls to 511, test to 77 with only 5 positives; ΔAP = 0.0, both λ = 0.001 — underpowered but not favorable to H1; 2,996/7,972 linked-article rows lack published_at and cannot be classified. (3) Alternate thresholds: ≥1 → ΔAP −0.0045; ≥3 → ΔAP −0.0066. Consistent. (4) Prevalence-standardized log loss (train prevalence): base 0.3278 vs aug 0.3323, Δ(base−aug) = −0.0045 — augmented still worse; the label-shift regime does not explain the null. (5) ±30-minute snapshot windows: not executable — the frozen-input amendment designates the extract as sole input; re-extracting from the drifted live DB would violate the frozen rule. (6) Matched-lambda sensitivity (reviewer-demanded): augmented at baseline-selected λ = 0.01 gives ΔAP = +0.0006; baseline at augmented-selected λ = 0.001 gives ΔAP = −0.0281; prevalence-standardized selection chooses λ = 0.1 for both models — like-for-like ΔAP = −0.0058 (the cross-rule +0.0074 is a selection artifact, not an effect of A). A matched-lambda permutation control (fixed λ = 0.01, 1,000 replicates, seed 20260817) reproduces the primary control's failure: observed ΔAP = +0.0006 vs null 95th percentile +0.0097. The null does not depend on the asymmetric regularization path. (7) Backfill within the primary test cohort: 77/115 test clusters contain at least one >24h-backfilled article (including all 30 positives); 304 of 686 linked-article rows through 18h are backfilled by this definition. The primary cohort is not backfill-free, but the exclusion robustness is consistent with the null.

Descriptive observations (no confirmatory status). Both models are well above chance: the baseline alone achieves AP 0.80 / AUC 0.897 — the outcome is learnable; only the incremental feature fails to add. The augmented fit gives A a positive standardized coefficient (+0.532) while the model shifts λ selection from 0.01 (baseline) to 0.001 and rebalances the trajectory features: backward title-arrival velocity flips sign (+0.448 → −0.345), the unique-title-count coefficient nearly doubles (+1.596 → +2.923), the article-count coefficient moves −0.078 → −0.596; stored heat velocity stays small (−0.019 → −0.038; no sign flip). On the 649 train+validation fitting rows, corr(A, V) = 0.216, A regressed on the eight baseline features has R² ≈ 0.657, and the early cumulative rate N3/a3 has std ≈ 1.42 — so A carries substantial variance beyond the baseline span, and the null is not a near-degeneracy artifact. Calibration is poor and similar in both models (intercept ≈ 0.52, slope ≈ 0.16): predicted probabilities are systematically overconfident; calibration is descriptive only. Warned category descriptives (future metadata): GEOPOLITICS n=50 ΔAP +0.011; CRYPTO n=21 ΔAP −0.082; STOCKS n=25 ΔAP −0.067; ECONOMICS n=19 ΔAP 0.0. None can alter the primary verdict; category is UNIDENTIFIABLE by design.

Discussion#

Robust findings. (1) A simple level/heat/velocity logistic model strongly discriminates future coverage growth at the 6h→18h horizon in this system (AP ≈ 0.80, AUC ≈ 0.897 on an untouched chronological holdout). (2) Adding the backward-only acceleration proxy does not improve — and slightly degrades — every evaluated metric; the effect is robust across bootstrap schemes, thresholds, prevalence standardization, and removal of the heat-velocity feature. (3) The pre-registered permutation negative control fails: the observed ΔAP is inside (below the 95th percentile of) the distribution of ΔAP values produced by permuted, information-free versions of A. The incremental-signal claim is falsified for this sample.

Preliminary findings. The fitted coefficients rebalance under augmentation — A receives a positive coefficient (+0.532), backward title-arrival velocity flips sign (+0.448 → −0.345), and the unique-title-count coefficient nearly doubles — while holdout performance does not improve. This is consistent with collinearity between A and the baseline trajectory features rather than a clean independent effect; it is an exploratory observation from one fit, not a tested mechanism.

Not-supported findings. H1, H2, H3, and both H4 thresholds are not supported. The hypothesized second-wave acceleration signal does not manifest in this production system's title-arrival records over this window.

Interpretation. The null is informative: if a system already tracks count, heat, and recent arrival velocity at hour 6, computing a two-interval acceleration proxy buys no measurable discrimination for the next 12 hours of coverage growth. This is a scoped statement about one clustering system, one 81.6-day window, one outcome definition, and one proxy; it does not license claims about other systems, horizons, or richer acceleration estimators (e.g., three-or-more-point curvature or point-process likelihoods).

Limitations#

1. Single system, single window. All conclusions concern one production clustering system's records over 2026-03-21→2026-06-11 (81.6 days). Snapshot cadence, clustering fingerprints, and deduplication rules are system-specific; results should not be generalized to other aggregators or venues without re-validation.
2. Proxy, not a derivative. A is computed from only two irregularly timed count intervals and is labeled UNRELIABLE as a measurement of true title-arrival acceleration. A richer estimator (3+ points, model-based rates) could in principle behave differently; the protocol tests the cheap operational proxy, not the continuous-time construct.
3. Missingness and backfill. Only 764 of 3,621 clusters with history qualified for all three ±1h windows; 2,996/7,972 linked-article rows lack publication timestamps, so backfill exclusion cannot be fully classified. Within the primary test cohort, 77/115 clusters (30 of 30 positives) contain at least one article added by 18h whose published_at precedes added_at by more than 24h. The backfill-excluded cohort (511 clusters) retains only 5 test positives and cannot adjudicate the null.
4. Outcome scope. The outcome is coverage growth as observed by the clustering system — not readership, public attention, causal virality, or price impact. Cluster merges/splits and title-fingerprint errors can contaminate lifecycle measurements, and rows within a cluster are dependent (mitigated by cluster-level resampling).
5. ±30-minute window robustness could not be executed under the frozen-input rule; the ±1h nearest-snapshot windows remain the only evaluated selection.
6. Label shift. Positive prevalence differs materially across partitions (23.6% / 61.7% / 26.1%); chronological evaluation and prevalence-standardized loss mitigate but cannot eliminate regime effects.
7. Calibration is poor in both models; probability magnitudes should not be treated as well-calibrated event probabilities.
8. Category effects are unidentifiable because category is not historized; all category tables are descriptive and future-contaminated.
9. Resolution relative to the materiality gate. The paired moving-block 95% CI half-width (~0.033 AP points) exceeds the pre-registered 0.02 materiality threshold, so this sample cannot reliably detect the study's own material effect size. The null is valid for the effect sizes this design can resolve; a smaller material threshold would require more test clusters.
10. Status. This is a preliminary pre-registered study; it has not been externally peer reviewed. All public claims are limited to the claims ledger, and the internal review process is described accurately.

Conclusion#

In the production market-news clustering system studied here, backward-looking unique-title arrival acceleration observed by hour 6 adds no held-out forecasting value beyond current title count, article count, heat, title-arrival velocity, heat velocity, and snapshot cadence for predicting a second wave of coverage (≥2 additional unique titles) over hours 6–18. The primary hypothesis is NOT_SUPPORTED under a pre-registered decision table: the point estimate is negative (ΔAP = −0.0121), the paired bootstrap interval straddles zero, all secondary tests are non-significant, and the permutation negative control fails. The baseline itself is strongly predictive, so the null is a genuine property of the acceleration proxy, not a failure of the data to be learnable. Teams operating similar ranking systems should not expect a two-interval acceleration feature to improve coverage-growth ranking once level, heat, and velocity are already conditioned on. The complete frozen protocol, executable artifact, byte-identical results, and provenance are stored in the run directory for independent replication. Explore all research

References#

1. Szabo, G., & Huberman, B. A. (2010). Predicting the popularity of online content. Communications of the ACM. https://arxiv.org/abs/0811.0405
2. Pinto, H., Almeida, J. M., & Gonçalves, M. A. (2013). Using early view patterns to predict the popularity of YouTube videos. WSDM 2013. https://doi.org/10.1145/2433396.2433453
3. Zhao, Q., Erdogdu, M. A., He, H. Y., Rajaraman, A., & Leskovec, J. (2015). SEISMIC: A self-exciting point process model for predicting tweet popularity. KDD 2015. https://arxiv.org/abs/1506.02594
4. Rizoiu, M.-A., et al. (2017). Expecting to be HIP: Hawkes intensity processes for social media popularity. WWW 2017. https://doi.org/10.1145/3038912.3052650
5. (2022). Hawkes processes for understanding heterogeneity in information propagation on social media. Frontiers in Physics. https://doi.org/10.3389/fphy.2022.1019380
6. (2024). Bayesian neural Hawkes process for event uncertainty prediction. International Journal of Data Science and Analytics. https://doi.org/10.1007/s41060-023-00443-3
7. Kleinberg, J. (2002). Bursty and hierarchical structure in streams. KDD 2002. https://doi.org/10.1145/775047.775061
8. Crane, R., & Sornette, D. (2008). Robust dynamic classes revealed by measuring the response function of a social system. PNAS. https://doi.org/10.1073/pnas.0803685105
9. (2022). A framework to predict early news popularity using deep temporal context networks. Expert Systems with Applications. https://doi.org/10.1016/j.eswa.2021.116060
10. (2025). SMTPD: A new benchmark for temporal prediction of social media popularity. https://arxiv.org/abs/2503.04446
11. (2025). Drift-aware neural Hawkes process for time-to-event modelling. WWW 2025 Companion. https://doi.org/10.1145/3703323.3703753
12. Ratkiewicz, J., Fortunato, S., Flammini, A., Menczer, F., & Vespignani, A. (2010). Characterizing and modeling the dynamics of online popularity. Physical Review Letters 105, 158701. https://doi.org/10.1103/PhysRevLett.105.158701

Related research: AI-Driven News Arbitrage in Perpetual Markets · Why Embedding-Based Sentiment Beats Traditional NLP for Crypto Trading