Multidimensional contemporaneous prediction of harmonized consumer price index levels across countries using reproducible machine learning and deep learning models

Accurate modeling of the consumer price index (CPI) is important for inflation monitoring and policy analysis. This study develops a multidimensional cross-country framework for CPI prediction and, in response to reproducibility concerns, replaces the principal neural-architecture ranking with a fully re-estimated and documented benchmark. The reproducible benchmark uses the supplied 545-observation, 39-country common-feature panel for 1995–2020, partitioned chronologically into 337 training observations (through 2014), 102 validation observations (2015–2017), and 106 holdout observations (2018–2020). For each country-year, CNN, RNN, and LSTM models receive up to the four most recent observed predictor vectors ending in year t; CPI itself is excluded from the input. Scaling is fitted on the training period only. Final architectures, selected hyperparameters, trainable-parameter counts, software versions, five fixed random seeds (101, 202, 303, 404, 505), per-seed holdout predictions, selected checkpoints, learning curves, and three expanding-window fold forecasts are reported and archived. The CNN contains 2,433 trainable parameters, the RNN 641, and the LSTM 7,073. Across five seeds, CNN has the lowest mean RMSE (0.1943, SD = 0.0094) and ranks first by RMSE in all five runs; LSTM has the lowest mean MAE (0.1128) and MAPE (0.0803). CNN also has the lowest RMSE in each of the three 2018–2020 expanding-window folds. Because the original TFT, GNN, and hybrid implementations cannot be reconstructed from the archived materials, they are removed from the principal comparative conclusions rather than assigned reconstructed architectures. The full-feature CatBoost/SHAP analysis and the lagged ex-ante benchmark branch remain separate analyses with their original information-set qualifications. The resulting conclusions emphasize reproducible performance, metric-specific trade-offs, and model-ranking stability rather than unsupported architecture superiority. Any legacy TFT, GNN, or hybrid point estimates that remain solely for traceability are explicitly treated as archived exploratory results and are not used for reproducibility claims, model rankings, inferential comparisons, or substantive conclusions. The chronological split is instantiated before every sample-dependent fitting operation, and the fitted StandardScaler objects, target-normalization parameters, split-membership files, and sequence-construction audits are preserved for the main split and for each expanding-window fold. No sample-dependent variable screening, PCA, graph construction, wavelet transformation, or imputation is used in this confirmatory principal benchmark; historical analyses requiring unavailable preprocessing objects remain outside the confirmatory conclusions. Because the exact row-level inclusion key and complete variable matrix for the historical 396-observation full-feature branch are not available in the current archive, that branch is retained only as archived exploratory/descriptive evidence and is excluded from the principal confirmatory conclusions; any interpretation of its SHAP, ablation, expanded-education, or subgroup results is limited to the data-observable country-years represented in that selected complete-case sample. Dependence-aware resampling further narrows the inferential claim: country-block and two-way crossed-cluster bootstrap intervals do not separate the reproduced neural architectures from one another under the loss functions associated with their reported rankings, whereas the CNN-CatBoost squared-loss difference remains below zero under both resampling schemes. The central empirical claims are restricted to the fully reproducible 545-observation benchmark: the final analysis dataset, executable estimation code, split-specific fitted preprocessing objects, fixed random seeds, observation-level predictions, learning curves, expanding-window fold forecasts, and dependence-aware comparison inputs are supplied as a single replication package. Analyses for which the complete row-level design matrix or implementation-defining objects are unavailable are excluded from confirmatory conclusions.