The academic document: introduction, theoretical framework, product, CRISP-DM methodology, tables and reproducibility.

1.1 Background

The Immigration and Nationality Act (INA) of 1965 [10] establishes the current system of annual quotas by category with a statutory 7 % limit per country (per-country limit), which is the primary source of the severe retrogression observed in high-demand countries such as Mexico, India, China, and the Philippines [6].

The Department of State publishes the Visa Bulletin [7] monthly with two calendars: Final Action Dates (FAD, published since 1992; the pipeline's homogeneous monthly series spans Dec-2001–present, 296 obs), which authorizes final adjudication, and Dates for Filing (DFF, since October 2015 with 130 obs), which authorizes the early start of the adjustment process. Bier [60] documents that the employment-based categories of countries under an effective limit, particularly India, have experienced retrogressions of several years in recent periods.

Prior work: Vegesana [11] applies discriminative classifiers to approval forecasting; Jain et al. [12] propose an ARIMA-LSTM hybrid for series with complex dynamics; Carammia et al. [13] and Pu et al. [9] integrate machine learning with data at scale in migration flows.

1.2 Problem definition

No published academic work was identified, in the preliminary review, offering an open, reproducible, and systematically evaluated predictive system over the Visa Bulletin multiseries panel with 95 % prediction intervals. Existing commercial platforms are black boxes with no methodology, data, or evaluation protocol published in an auditable way.

The direct consequence is that millions of applicants must plan medium- and long-term decisions (family, professional, and financial stability) on opaque estimates or low-quality methods. Backlog figures are documented in Section 1.1 [4], [6].

The need is to build a public longitudinal database of the Visa Bulletin 1992–2026 and an evaluable predictive system that turns more than three decades of data into auditable forecasts with 95 % prediction intervals, filling a verifiable gap in the applied literature.

1.3 Objectives

General objective

Develop and implement an applied predictive system for Visa Bulletin priority dates, organized as a multiseries panel $y_{p,c,b,t}$ indexed by chargeability country, immigrant category, and table type, with monthly forecasts at horizons of 1, 3, 6, and 12 months accompanied by 95 % prediction intervals, under the CRISP-DM methodology [64].

Specific objectives

1. Build the longitudinal database of the Visa Bulletin 1992–2026 structured by country × category × table × month, published under an open license.
2. Empirically characterize the historical behavior of priority dates and their administrative regimes by chargeability country or area.
3. Design and implement the full family of models listed in Table 3, without privileging any architecture in advance.
4. Evaluate predictive performance through expanding walk-forward temporal validation with scaled metrics (sMAPE, MASE, MAE, RMSE) and 95 % prediction intervals, reporting aggregate and disaggregated results by country, category, and table type.
5. Deliver an academic demonstration application that allows querying historical series, forecasts, and 95 % prediction intervals with explicit warnings about the informational and non-legal nature of the estimates.

1.4 Justification

Social contribution

The process of obtaining permanent residence affects millions of people in terms of family, work, and financial stability. Nearly 4 million remain abroad awaiting a family-based visa [6], within the global USCIS backlog of 11.5 million pending cases across all forms [4]. The project supports applicants' planning with 95 % prediction intervals and greater transparency than the closed, non-auditable tools on the market.

Technical contribution

The system addresses observable limitations of closed commercial platforms, particularly the lack of transparency regarding data, methodology, and evaluation protocol. The integration of a complete multiseries panel (multi-country × multi-category × multi-table) under a rigorous comparative framework constitutes an applied contribution to the forecasting of migratory phenomena.

Academic contribution

• Open database 1992–2026 with a permissive license, pre-registered evaluability criteria, and verifiable reproducibility (R1–R7, Appendix A.3).
• Named comparative framework under CRISP-DM with a family of linear and nonlinear models evaluated without privileging architectures in advance.
• Uncertainty quantification with three mechanisms for the 95 % prediction intervals: analytical ARIMA, MC dropout [40], and conformal prediction [59].

1.5 Questions and hypotheses

Consistent with the professionalizing profile of the MIAAD program, the questions are framed as open questions and the hypotheses as soft qualitative statements that will be tested empirically against the pilot data. Formal statistical rigor (Diebold-Mariano + Holm correction at $\alpha =0.05$) is documented as a methodological decision in Chapter IV, not as a pre-committed threshold in the body.

General question

With what accuracy and with what calibration of the 95 % prediction intervals can the future behavior of the Visa Bulletin multiseries panel be forecast under a CRISP-DM comparative framework, over the coverage of evaluable series and respecting the structural heterogeneity among country–category–table cells?

Specific questions

1. Patterns. What patterns of advancement, stagnation, and retrogression do the series present by chargeability country and immigrant category, and what proportion of each series is evaluable under the pre-registered criteria?
2. Predictive performance. Which of the implemented models —linear (ARIMA, SARIMA, Prophet) or nonlinear (LSTM, ARIMA-LSTM, DeepAR, XGBoost)— offers the best predictive performance in sMAPE and MASE over the evaluable pilot set, and with what consistency across strata?
3. Series characteristics. What structural characteristics (effective length, frequency of retrogressions, frequency of Current/Unavailable states) explain predictive difficulty per cell?
4. Coverage. What empirical coverage do the 95 % prediction intervals achieve relative to their nominal level, in aggregate and by stratum?

Hypotheses

These hypotheses are interpreted in terms of predictive associations that are empirically verifiable with the project's data, not as causal claims about the administrative dynamics of the Department of State. Causal interpretation is explicitly out of scope.

3.1 Description of the solution

3.1.1 · Methodological framework

A comparative framework that does not privilege architectures in advance, instrumented under CRISP-DM [64]. Three complementary families: (a) linear models (seasonal naïve, ARIMA, SARIMA, Prophet), (b) local nonlinear models (LSTM, ARIMA-LSTM), (c) global and tabular models (DeepAR, XGBoost). The identity of the best-performing model is an empirical question of the study.

3.1.2 · Tentative architecture

Note (July 2026): the figure is the literal tentative architecture from the May 2026 proposal, whose comparison framework proposed that initial set of candidate models. The framework actually executed grew during development — the current count is derived from the governed model catalog and is reported in the prospective evaluation.

3.1.3 · Analytical coverage at three levels

The system's coverage is organized into three explicit levels:

• Structural database. Contains all combinations of country (Mexico, India, China, Philippines, All Chargeability Areas Except Those Listed) × category (family-based F1, F2A, F2B, F3, F4 and employment-based EB-1 through EB-5 with subcategories) × table (FAD, DFF).
• Evaluable coverage. Series with sufficient history, continuity, and variability for forecasting, according to the pre-registered criteria in Table 2.
• Initial pilot coverage. Family-based categories F1–F4 over the high-demand countries and the residual grouping, starting with Mexico because of the severity of its backlogs.

3.1.4 · Predicted variable

Let $y_{p,c,b,t}$ be the priority date transformed into days since a base date, for the cell (country $p$, category $c$, table $b$) in month $t$. The variable is continuous and constitutes the target of a single temporal regressor. The system is trained exclusively on observations with state $e=F$ (specific date); Current and Unavailable observations are preserved in the database as descriptive annotation but are not a predictive target. Retrogressions within the $F$ regime are kept as legitimate observations (a genuine phenomenon of the immigration system).

3.1.5 · Training isolation

The regression metrics are computed exclusively on observations with a specific date. This avoids contaminating the evaluation with categorical observations (Current/Unavailable) that fall outside the system's predictive target. The descriptive characterization of the C/U/F regimes lives in the exploration report (Ch. IV, Phase 2), not in the operational metrics.

3.2 Delimitations, limitations, and governance

The scope is bounded through an explicit set of delimitations (sovereign decisions of the researcher), while the constraints imposed by external factors are acknowledged as limitations.

Scope governance

3.3 Form of validation

Chapter III describes what is validated and at what level success is reported. The operational decisions (walk-forward parameters, confirmatory testing scheme, family-wise error control) are documented in Chapter IV.

Success levels