CPIIW Forecast 2024–2025: How Machine Learning Predicts India's Industrial Worker Inflation

Data and preprocessing

The study uses monthly CPIIW data (base 2016 = 100) from the Labour Bureau, Chandigarh, spanning January 2006 to December 2023 — 216 observations. Before modeling, the authors ran three standard diagnostic tests. A Shapiro-Wilk test confirmed the data isn't normally distributed. A Ljung-Box test showed the series is not random white noise (p ≈ 0.00), confirming genuine autocorrelation rather than randomness. An Augmented Dickey-Fuller (ADF) test for stationarity returned a p-value of 0.98 — far above the threshold needed to reject non-stationarity — confirming the raw series is non-stationary: its statistical properties shift over time. First-order differencing was applied to stabilize the series before modeling.

Modeling pipeline

Twenty-seven models were fit using the PyCaret time-series library in Python, spanning simple baselines (naive, seasonal-naive, grand-means forecasters), classical statistical methods (ARIMA, exponential smoothing, Theta forecaster), and regression-based approaches adapted for time series with deseasonalizing and detrending preprocessing (linear regression, Huber regression, ridge, lasso, and related variants). Separately, the authors built an LSTM (Long Short-Term Memory) neural network — an advanced type of recurrent neural network designed to remember long-range patterns in sequential data — and evaluated it on the same held-out data using the same four metrics: MAE, RMSE, MAPE, and R².

The winning model: LSTM

A plain recurrent neural network struggles to learn dependencies that stretch far back in a sequence — the influence of early inputs tends to vanish as it propagates forward. LSTM solves this with a gating mechanism that explicitly controls what a memory cell keeps, discards, and adds at every time step, following the standard formulation:

Model performance: R² across the field

R² measures the share of variation in CPIIW a model explains. LSTM's 0.992 was in a different tier entirely from the rest of the field — the next-best model, a Huber regression with deseasonalizing and detrending, trailed by ten percentage points, and the gap to simple baseline forecasters was far larger still.

Generalization: training vs. test performance

A model that fits training data well but falls apart on new data isn't actually useful for forecasting. LSTM's performance barely changed between the training set (R² = 0.9921, MAPE = 0.52%) and the held-out test set (R² = 0.9918, MAPE = 0.50%) — a sign the model generalizes rather than simply memorizing the historical series.

The two-year CPIIW forecast

Applying the trained LSTM model, the study projects CPIIW forward month by month from January 2024 through December 2025. The seasonal shape is the notable part: the index stays nearly flat over winter (October–March) but rises by roughly 3.4 points during the summer/monsoon stretch (April–September) each year. Food, Beverages and Tobacco carry nearly half the basket's weight, and monsoon-season rainfall variability — too little rain hurting crop output, too much damaging standing produce — tends to push food prices up sharply in exactly that window.

Dearness Allowance implications

DA/DR for Central Government employees and pensioners is calculated from CPIIW movements under the formula set by the 7th Pay Commission, so the CPIIW forecast translates directly into an estimated DA/DR trajectory.

A companion result: the same method applied to the US

In the same journal issue, the same author team (Kambo and Jassimar Singh, joined here by Gurinder Singh) ran a directly comparable exercise on US CPI-U inflation, testing 27 models. There, a classical statistical method — ETS (Error, Trend, Seasonality) — won instead of LSTM, not any machine-learning model. That a different model wins on India's series than on the US series is itself informative: there's no universally best forecasting method for a consumer price index, and model choice needs to be validated separately for each country's data rather than assumed to transfer.

Limitations

This is a statistical projection based on historical patterns in the CPIIW series itself, not an economic model incorporating policy decisions, wage-commission revisions, or external shocks — an unusually severe monsoon failure or a sudden food-price spike, for instance, would move CPIIW in ways a purely time-series model, however well-fitted to the past, would not have anticipated in advance. The DA/DR figures above are the study's estimates derived from the CPIIW forecast via the 7th Pay Commission formula, not official rates announced by the Government of India, which are set separately each January and July.