Cahiers de recherche UQO · WP39 · v1.0
The Invisible Balance Sheet
Can Intraday Financial Markets Detect Corporate Deterioration Before Accounting Disclosures?
- Année
- 2026
- Première parution
- novembre 2026
- Dernière révision
- novembre 2026
- Statut de publication
- Cahier de recherche
The invisible balance sheet is visible in the daily close; the minute-by-minute tape adds noise, not foresight
On 53,667 untouched 2021–2026 firm-quarters, adding the same realized-measure, volume-composition and co-movement families computed from 1-minute bars to a point-in-time fundamentals-plus-daily benchmark changes the out-of-sample log loss by +0.0013 [−0.0013, +0.0041] (logistic) and +0.0016 [+0.0010, +0.0023] (gradient-boosted trees), against a daily increment of −0.0068; the null holds at five horizons, for five accounting targets, a volatility placebo, unseen firms, every refit year and every subgroup.
Résumé
Do one-minute financial markets see a firm's forthcoming accounting deterioration before daily markets do? This paper builds a strictly point-in-time panel of 3,796 U.S. listed firms and 102,522 firm-quarters (2013–2026) from 6.7 billion raw one-minute bars and quarterly statements with filing and release dates, and forecasts an adverse operating-income surprise, defined against a seasonal random walk with drift and never against prices, twenty trading days before the earnings release. Holding the information set, the universe, the horizon and the model capacity fixed, the same families of realized measures, volume composition and market co-movement computed from 30-minute, 5-minute and 1-minute bars are added to a benchmark that already contains point-in-time fundamentals and daily OHLCV, with an identical 24-trial tuning budget for every information set and twelve leakage guards implemented as unit tests.
Fundamentals explain most of the predictable variation (log loss from 0.6282 at the base rate to 0.4937 with gradient-boosted trees, AUC 0.797), daily market data add a clear increment (−0.0068 [−0.0095, −0.0042], AUC 0.803), and one-minute data add nothing: the paired out-of-sample change in log loss on the untouched 2021–2026 test period of 53,667 firm-quarters is +0.0013 [−0.0013, +0.0041] with logistic regression and +0.0016 [+0.0010, +0.0023] with trees, intervals that exclude any gain at all with trees and any gain larger than a fifth of the daily increment with logistic regression, against a pre-registered minimum detectable difference of 0.0025. The 30-minute and 5-minute families behave identically (+0.0017 and +0.0021 with trees), and concatenating all three resolutions is worse still (+0.0030).
The null survives five horizons from forty days to one day before the release (rank correlation between gain and proximity −0.50, p = 0.39), five accounting targets, a symmetric volatility placebo, firms never seen in training (+0.0018 on held-out firms), every annual refit and every pre-specified subgroup. The small information content of the incremental one-minute signal is concentrated in large, levered firms rather than in neglected ones (panel coefficient −0.0129, t = −1.76, against −0.0765, t = −32.22, for the daily benchmark signal), is already impounded in prices at the cutoff (announcement-window abnormal return unrelated to the signal, t = 0.01), and has no economic value in factor-adjusted portfolio sorts (FF5+momentum long–short alpha −0.59% per month, t = −1.85, gross; −0.62% net of 20 bps per side). The accounting panel is conditioned on firms listed in November 2026, a survivorship caveat the paper documents rather than resolves.
Résultats clés
3,796 firms · 102,522 firm-quarters
strictly point-in-time panel (2013–2026) built from 6.7 billion raw 1-minute bars for 16,195 U.S. symbols and quarterly statements with filing and release dates
−0.0068 [−0.0095, −0.0042]
daily-market increment in out-of-sample log loss over point-in-time fundamentals (LightGBM, AUC 0.797 → 0.803): prices lead earnings, reproduced
+0.0016 [+0.0010, +0.0023]
1-minute increment with gradient-boosted trees on 53,667 test firm-quarters: a small deterioration, excluding any gain at all (MDE 0.0025)
+0.0013 [−0.0013, +0.0041]
1-minute increment with logistic regression (Giacomini–White p = 0.518): no gain larger than a fifth of the daily increment
ρ = −0.50, p = 0.39
no horizon gradient from 40 to 1 trading days before the release; placebo AUC 0.503 on shuffled labels; held-out firms +0.0018
−0.59 %/month (t = −1.85)
FF5+momentum alpha of long–short decile sorts on the 1-minute score, −0.62 % net of 20 bps: no economic value; announcement-window CAR unrelated to the signal (t = 0.01)
Contributions scientifiques
- A frequency-controlled test of market anticipation
- Separates anticipation (prices move before statements, expected under semi-strong efficiency and reproduced) from the marginal information value of sampling frequency, with a nested ladder of information sets B1–B5 / LB1–LB5 sharing the same universe, calendar, cutoff, horizon, feature families and tuning budget, so that any difference is attributable to resolution rather than to capacity.
- A precise null, not an underpowered one
- The pre-registered minimum detectable difference at 80 % power (0.0025 log-loss units) is wider than the boosted-tree intervals and comparable to the logistic ones; the 1-minute increment is zero within tight bounds while the daily increment (−0.0068) is large and unambiguous, and 30-minute and 5-minute families behave identically.
- Five pre-registered hypotheses, all answered
- H1 (incremental information) and H2 (resolution ordering) rejected; H3 (heterogeneity) reversed in sign at zero level, with the little incremental content sitting in large, levered firms rather than small, illiquid ones; H4 (timing) rejected at five horizons and in the fixed-calendar design; H5 (pricing) rejected on announcement-window returns and factor-adjusted decile portfolios.
- Strict point-in-time discipline on market and accounting clocks
- The reveal date is the earnings press release, not the filing; the information set at cutoff c holds bars with session ≤ c and facts with available_at ≤ c only; twelve leakage guards (future facts, fiscal-end joins, post-cutoff bars, test-period transforms, daylight-saving errors, identifier drift, split artefacts, release-day bars, overlapping windows, survivorship, ingestion dates, randomised placebos) are unit tests, and the local as-of join is verified live against the API.
- Bounds on the high-frequency mechanisms
- Whatever informed trading, jump arrival or gradual information processing is visible in one-minute bars twenty days before a release is summarised by the daily close, paralleling for fundamentals what the variance-measurement literature established: the finest grid is not the most informative.
- Two structural limits stated, not hidden
- The accounting panel is survivor-conditioned (firms listed in November 2026) and the vendor's point-in-time stamps (filing date + 1 day, revision 1 values) are a documented proxy; neither affects the cross-frequency comparison but both affect the level of predictability, and three internal referee reports list the v1.1 revisions.
Méthodologie
Méthodologie (5)
Event-time clocks and information set
Cutoff c = A_q − h trading days for h ∈ {40, 20, 10, 5, 1} (h = 20 primary), admissible only if c < min(A_q, F_q) and c > A_{q−1}; a fixed-calendar design with month-end cutoffs serves the pricing tests and as a check that does not condition on the release schedule.
Accounting surprise targets
Scaled operating-income surprise against a seasonal random walk with drift computed from quarters available at c, with the adverse event defined as a shortfall of at least 0.5 % of assets; secondary net-income, ROA, operating-cash-flow and liquidity targets, a symmetric |S| ≥ 0.5 % placebo and a ridge model of the continuous surprise.
Same 19 feature families at three resolutions
Realized variance, bipower variation, relative jump, signed semivariances, realized skewness and kurtosis, concentration, autocorrelation, intraday drift, opening and closing volume shares, volume Gini, down-bar volume share, bar presence, realized beta to SPY, market-down variance share, variance ratio and volatility-of-volatility, computed from 30-minute, 5-minute and 1-minute bars over the 20-session window; 16 daily and 21 point-in-time accounting columns in the benchmark.
Equal-capacity ladder and chronological validation
ℓ2 logistic regressions tuned on the validation period and LightGBM with an identical 24-trial random-search budget per information set; training on 2012–2018 cutoffs, tuning on 2019–2020, testing on 2021–2026 with purging and a one-quarter embargo; expanding-window annual refits from 2016 and a 20 % firm hold-out.
Inference, mechanisms and pricing
Block bootstrap over calendar quarters of the reveal date (1,000 draws), Giacomini–White tests with Newey–West variance and Holm correction within hypothesis families; shuffled-label negative control; panel regressions of the realised surprise on out-of-fold deterioration signals with firm and sector-by-quarter fixed effects and two-way clustering; announcement-window CAR regressions and monthly decile sorts with FF5+momentum alphas net of 20 bps per side.
Données de recherche
- FinData OS lakehouse of FirstRate Data
- Private lakehouse operated by the author and accessed through its versioned API: 6.7 billion raw 1-minute bars for 16,195 U.S. stock symbols (2000–2026, including 8,069 delisted symbols), quarterly fundamentals for 5,451 entities (67 metrics, filing dates 2012–2026) and 373,326 earnings-release events with their timing; no raw record is redistributed.
- Universe and event-time sample
- 4,568 firms with at least eight quarterly periods, 250 sessions of 1-minute bars after 2012 and a non-governmental SIC division (5,073 matched to a CIK); 194,817 firm-quarters, 481,159 event-time observations across five horizons, a primary h = 20 cohort of 97,788 observations with cutoffs from 2013-09-26 to 2026-10-01 and a 30.1 % adverse rate; 88.5 % of firm-quarters matched to a release with a filing lag of at most 30 days.
- Public augmentations
- SEC company_tickers.json and submissions files for CIK and SIC industry, and the Kenneth French daily factor library for the FF5+momentum alphas of the portfolio tests.
Reproductibilité
- Pipeline scripts 00–09 run on the FinData storage node against the local versioned API; hashed extraction manifests (dataset identifiers and versions, row counts, SHA-256 digests) pin every input and the design freeze hash is recorded before the test period is opened
- Hypotheses, endpoints, horizons and decision rules frozen in docs/RESEARCH_DESIGN.md; twelve leakage guards are unit tests, including a live equivalence test of the as-of join against the FinData API's as_of semantics
- FirstRate Data is licensed for private research use: no raw record is redistributed, the package contains code, manifests and derived aggregates only
- Three internal referee reports with author responses list the v1.1 sensitivity analyses (availability shift, restatement bound, noise-column control, short window); full LaTeX source of the 34-page paper with 206 auto-generated result macros
Artefacts de recherche
Historique des versions
| Version | Date | Description | Fichier |
|---|---|---|---|
| v1.0courante | novembre 2026 | Manuscrit courant tel que publié sur www.spboucher.ai (édition révisée après la révision scientifique du 2026-10-17 pour les cahiers 1–17). | PDF · 34 p. |