DSA210 / Methods

Keep the future out of training.

A reproducible, retrospective design. The same calendar boundary applies to every country, and every preprocessing step learns only from the training window.

The analysis protocol

01

Define consumer confidence

Official OECD Consumer Confidence: MEASURE=CCICP, ADJUSTMENT=AA, TRANSFORMATION=IX, UNIT_MEASURE=IX, dataset version 4.1.

02

Prepare the research panel

Use population-weighted weather and macro/retail snapshots. Preserve missing values, verify unique country-month keys and contiguous months. Do not impute any target.

03

Split by calendar, not by row

For test year Y, train on dates strictly before December of Y−1. Test on January–December of Y. This leaves a one-month embargo. All countries share the split, so a month cannot be both train and test.

04

Fit transformations inside each fold

Median imputation, missingness indicators, scaling, one-hot country encoding and weather PCA are fitted only on training rows. PCA uses two components. Report retained variance separately for each fold.

05

Compare fixed, interpretable models

Use country-mean and last-observed-month persistence baselines, three matched feature-group forests, a forest with the lagged target, and Ridge with weather PCA. RF: 160 trees, depth 6, minimum leaf 5; Ridge alpha 10; no outcome-driven hyperparameter search.

06

Respect the next-month label

Classify whether CCI at t+1 exceeds CCI at t. Drop the four undefined terminal labels. Training label dates must precede the test boundary. Logistic settings are fixed; report AUC and accuracy against a training-prior baseline.

07

Report uncertainty and scope

Country-specific regression includes month fixed effects and a linear trend; HAC(12) errors, with small-sample correction. Weighted-city, capital-city and first-difference specifications form one 12-test BH family. No specification is selected after seeing its p-value.

Six windows into future periods

2019
Train through 2018-11 · test 2019-01—2019-12
2020
Train through 2019-11 · test 2020-01—2020-12
2021
Train through 2020-11 · test 2021-01—2021-12
2022
Train through 2021-11 · test 2022-01—2022-12
2023
Train through 2022-11 · test 2023-01—2023-12
2024
Train through 2023-11 · test 2024-01—2024-12

Green: training · empty month: embargo · orange: test · the full track is 2015–2024.

What this is—and what it can claim

These are retrospective, out-of-time evaluations with revised historical data. Same-month weather and macro observations may not be available at an operational forecast origin. Results are conditional explanatory benchmarks, not a certified real-time forecasting service. Persistence uses the previous observed month, including earlier observations within a test year.

Sources and reproduction

The downloadable panel, model predictions and source manifest provide the evidence behind the study. Research code ↗