A data-science investigation · 2026

Can we predict the heartbeat of Indian monetary policy one week ahead?

Ten years, 545 weeks, 117 features, two monetary-policy regimes, and one overnight rate — the WACMR — that quietly settles every bank's cash book. We combined NDAP RBI data with market prices and 75 curated policy events to build a forecasting model with 70.9% directional accuracy, explain it with SHAP, and let you interrogate it.

WACMR (cyan fill) vs Repo Rate (dashed). Amber bands are Regime 1 (Tightening era, 315 weeks); green bands are Regime 0 (Accommodation era, 230 weeks).
Why it matters

The WACMR is the thermometer of Indian monetary policy.

The Weighted Average Call Money Rate is the interest rate at which scheduled Indian banks lend to each other overnight, settled on the books of the Reserve Bank of India. It's the point where the RBI's policy rate, the banking system's liquidity state, and market expectations all have to agree.

If you want to know whether monetary policy is actually transmitting — whether the repo rate is biting — the WACMR is where you look. Forecasting it a week ahead is genuinely hard: the series lives inside a tight corridor, but with sharp, regime-dependent excursions around COVID, rate cycles, and liquidity operations.

This project frames the problem, pulls together open data from NDAP and elsewhere, engineers a feature set, clusters the dataset into market regimes, trains a walk-forward-validated XGBoost forecaster, opens the model with SHAP, and turns the whole thing into an interactive artefact you can play with.

Three findings that shaped everything else

What the data told us.

The rate corridor dominates
≈ 90%

Mean |SHAP| concentrates on the repo rate, its one-week lag, and the WACMR–repo spread. Equity and forex features do not appear in the top 15.

COVID was a genuine break
2 regimes

PCA+K-Means on 90%-variance components splits the sample into a pre-COVID tightening regime and a post-COVID accommodation regime at a silhouette-optimal k=2.

Walk-forward, not overfit
70.9%

Directional accuracy on one-week-ahead predictions using an expanding-window XGBoost with a 156-week minimum train set (RMSE 0.102, MAE 0.065).

The pipeline

Six stages, reproducibly.

01
Ingest

8 RBI datasets + Yahoo Finance, aligned to a Friday weekly grid

02
Engineer

Technical indicators, lags, spreads — 117 features

03
Regimes

PCA → K-Means (k=2) with silhouette validation

04
Forecast

XGBoost, expanding-window walk-forward

05
Explain

SHAP TreeExplainer per week + aggregate

06
Narrate

75 curated policy events with manual sentiment

Explore

Where to go next.

Built for the DSM project, 2026. Data from the NITI Aayog National Data & Analytics Platform and Yahoo Finance. Model, regimes, and SHAP via open-source packages. The research agent is powered by Gemini 2.5 Flash with function-calling.