MICAI 2026 · Paper #12 · Chihuahua, 2–6 November 2026
A framework for auditable per-series model selection for weekly reported depression cases in Mexico (ICD-10 F32, SINAVE), with its recorded 2026 forecasts evaluated retrospectively.
¹ Tecnológico de Monterrey · ² Instituto Mexicano del Seguro Social, OOAD Guerrero
Independent research; affiliations do not imply institutional endorsement.
Interactive follow-up
Pick any of the 99 targets. The chart shows the recorded path of the assigned engine, a copy of the value 52 weeks earlier and the reported cases up to the last period in this data snapshot (W38, edition W39).
Recorded forecasts · retrospective evaluation
In the talk, the national paper window uses editions W02–W18: Ensemble 7.40% vs. naive 7.33%.
| Estimator | A | B | C |
|---|---|---|---|
| Ensemble (assigned) | 7.28% | 5.64% | 6.35% |
| 52-week naive | 7.41% | 6.36% | 6.82% |
| Prophet | 8.36% | 6.22% | 7.15% |
| DeepAR | 10.83% | 7.52% | 8.95% |
| Stacking | 16.26% | 22.98% | 20.08% |
Period W01 reproduces the paper; its legacy join is ambiguous. Prophet, DeepAR and Stacking start at W02. The talk uses editions W02–W18 (7.40% / 7.33%); table A uses W03–W18 (7.28% / 7.41%).
What the record shows
Historical selection favored DeepAR, and the archived assignment picked it for all 96 state targets. Its historical median (5.7% against 26–31%) was not repeated: in the paper window, editions W03–W18, its recorded median was 27.4%. The historical state DeepAR code scores training fit; a common validation protocol is not verified.
Excluding DeepAR, Ensemble had lower historical sMAPE than Prophet in 61% of targets. In the paper window, editions W03–W18, Prophet was lower in 60%. These comparisons are post hoc, exploratory and based on overlapping targets.
In this seven-week replay (re-selecting with editions W03–W11, scoring W12–W18), revision reduced the median from 32.7% to 28.5%. Simple alternatives had lower medians: Prophet alone 27.6%, a pool without DeepAR 26.4%.
In the 21-week follow-up, bulletin editions W19–W39, the recorded national Ensemble path has 5.64% sMAPE against 6.36% for the naive. Across the 99 targets, the archived assignment has a higher median than the naive (31.7% vs. 28.6%) and a lower sMAPE in 60 of 99. Prophet and Ensemble beat the naive in 72 and 70 of 99. This motivates reassessing the assignment; revision in this later window remains untested.
| Window | Prophet | DeepAR | Ensemble | Stacking | Archived | Naive |
|---|---|---|---|---|---|---|
| A · W03–W18 | 26.7 (84/99) | 27.4 (73/99) | 28.1 (72/99) | 29.3 (58/99) | 27.4 (75/99) | 33.7 |
| B · W19–W39 | 24.2 (72/99) | 31.7 (57/99) | 25.7 (70/99) | 31.4 (48/99) | 31.7 (60/99) | 28.6 |
| B-sens · W19–W37 | 23.2 (72/99) | 30.8 (56/99) | 25.7 (66/99) | 31.2 (47/99) | 30.8 (59/99) | 28.4 |
| C · W03–W39 | 26.4 (77/99) | 30.1 (65/99) | 27.3 (72/99) | 32.0 (55/99) | 30.1 (67/99) | 30.1 |
The talk
10 minutes plus 5 for questions, in English. The Spanish version is a parallel copy for co-authors.
The paper
To appear in Advances in Soft Computing, MICAI 2026, Springer LNCS 17149. The accepted manuscript is available here; the Version of Record will be available from the publisher.
@inproceedings{rebull2026micai,
title = {Forecasting Weekly Depression Incidence in Mexico: A Multi-Model Framework with Auditable Per-Series Model Selection},
author = {Rebull-Saucedo, Javier Augusto and P{\'e}rez-Nava, Juan Carlos and S{\'a}nchez-Salazar, Luis Gerardo and Barcel{\'o}-Alonso, Grettel and P{\'e}rez-Hern{\'a}ndez, Ruth},
booktitle = {Advances in Soft Computing. MICAI 2026},
series = {Lecture Notes in Computer Science},
volume = {17149},
publisher = {Springer},
year = {2026},
note = {To appear}
}
Reproduce
Code, data and the released assignment (GitHub) · Data used on this page (JSON)