Bank Records, Reporting Systems, and Chronological Validation

Conceptual map

IV-17.01Field meaning
IV-17.02Posterior construction
IV-17.03Record acquisition
IV-17.04Prediction baselines
IV-17.05Calibration and validation

1. Database fields are outcomes of a reporting system

A bank record combines an economic event with definitions, reporting thresholds, release delays, revisions, and missing-data rules. Likelihoods must respect that observation process.

Let latent state $\theta$ generate an economic quantity $$X$$ , and let the reporting mechanism produce field $$Y$$ through kernel $Q(dy|x,\theta ,v)$ , where $$v$$ denotes the reporting vintage. Structural zeros, censored small values, missing records, and positive measurements are distinct events unless the documentation equates them. Entity identifiers, mergers, restatements, currency, units, and consolidation scope determine whether records can be compared through time.

Definition 1 · Chronological information set

$\mathcal{F}_{t}^{(v)}$ contains only records released by prediction date $$t$$ in vintage $$v$$ . A validation exercise is chronological when every fitted model and prediction uses the information set that would have been available at that date.

2. Point masses and densities require a common dominating measure

Consider a structural-zero exponential record: $$Y=0$$ with probability $\pi$ , and conditional on being positive, $$Y$$ has exponential rate $\lambda>0$ . With $\mu=\delta _{0}$ +Lebesgue measure on (0, $\infty$ ), the likelihood contribution is

\[p(y|\pi ,\lambda )=\pi 1\{y=0\}+(1-\pi )\lambda e^{-\lambda y}1\{y>0\}, \quad 0\le \pi \le 1.\](1)

Proposition 1 · Normalization of the mixed record law

For $0\le \pi \le1$ and $\lambda>0$ , (1) integrates to one with respect to $\mu$ . Its probability at exactly zero is $\pi$ , while its continuous component has total probability $1-\pi$ .

Proof. Integration over the atom gives $\pi$ . Integration over the positive half-line gives $(1-\pi)\int _{0}^{\infty}\lambda e^{-\lambda y}\,dy=1-\pi$ . The components sum to one and are nonnegative. ∎

At $\pi=0$ the model is purely exponential and assigns zero probability to the singleton {0}; at $\pi=1$ , $\lambda$ is unidentified because no positive observations occur. A zero caused by rounding or left censoring has probability obtained by integrating the continuous density over the relevant interval and should not be labeled a structural atom without evidence.

3. Updating calculations depend on the recorded event

With $\pi=0.3$ and $\lambda=2$ , $$P(Y=0)=0.3$$ . At positive value $$y=0.5$$ , the density with respect to Lebesgue measure is $0.7\times 2\times e^{-1}\approx0.515031$ . This density value is not a probability at the single point. Integrating the positive density yields 0.7.

For a separate Gaussian example, let $\theta \sim N(0,1)$ and observe $S=\theta+\varepsilon$ with $\varepsilon \sim N(0,1)$ independent. Completing the square gives $\theta|S=s\sim N(s/2,1/2)$ . If the signal is absent because a report was never released, inserting zero as $$s$$ imposes a false observation. The correct posterior either conditions on the missingness event or remains at the prior when missingness carries no state information.

4. Revisions and acquisition dates determine honest validation

Failure case · Final-vintage fields used in historical prediction

A backtest that trains or predicts with later revisions gives the historical model information unavailable at its forecast date. Random train-test splitting also lets future reporting regimes leak into earlier predictions.

A rolling-origin design fits on releases available through date $$t$$ and evaluates later outcomes, then advances the origin. Hyperparameter selection must be nested inside each training window or fixed in advance. Entity entry and exit alter the prediction population; evaluation must state whether the target is a fixed cohort, all active entities, or entrants. When labels mature with delay, the training cutoff follows label availability as well as covariate release. Empty windows, all-zero cells, $\pi$ boundaries, and sparse entity clusters need explicit handling. Calibration plots should use held-out forecasts and intervals should include parameter, reporting, and revision uncertainty when those are part of the target.

5. Implementation, exercises, and sources

Build an append-only acquisition table with source identifier, entity key, period covered, release timestamp, vintage, field definition, unit, value, missing code, and revision link. Construct analysis snapshots by an as-of join on the prediction date. Test mass and density normalization, support, limiting parameter cases, and posterior calculations before estimation. Preserve the untouched final evaluation window.

Download the volume verification script →

Exercises

Verify the density and total masses in Section 3.
Derive the Gaussian posterior by adding prior and signal precisions.
Write the likelihood when a latent exponential value is reported as zero whenever it is below threshold $$c>0$$ .

Partial solutions

1. The atom contributes 0.3 and the integrated density contributes 0.7; substitution at 0.5 gives 1.4/e. 2. Prior and signal precisions are one, so posterior precision is two, variance one half, and precision-weighted mean is $$s/2$$ . 3. The reported-zero probability is $1-e^{-\lambda c}$ ; positive reported values have density $\lambda e^{-\lambda y}$ on $y\ge c$ , subject to the declared convention at equality.

James O. Berger, Statistical Decision Theory and Bayesian Analysis.Likelihood construction and posterior decisions.
Diane Lambert (1992), “Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing,” Technometrics 34, 1–14.Mixture modeling with structural zeros.
Chae-Yeon Xon (2026), “Economica Obscura: Institutions, Archives, and Policy Learning.”Related archival reporting-system application.

6. Audit checkpoint

Latent event, field definition, atom versus censoring, dominating measure, parameter support, $\pi$ endpoints, positive-rate condition, missingness mechanism, release time, vintage, revisions, entity history, consolidation scope, units, as-of join, label maturity, rolling origin, nested tuning, cohort definition, calibration sample, empty and all-zero windows, and final holdout integrity.

7. Scope boundary

The chapter covers mixed likelihoods, elementary Gaussian updating, and chronological validation. Dynamic latent-state systems, informative acquisition policies, and regulatory reporting models require additional structure.