1. Why a forecast is not yet a policy
A model that predicts a number is a model. A model that changes how someone makes a decision is a tool. Most of the work this project showcased — the regimes, the SHAP attributions, the walk-forward residuals — sits cleanly on the model side of that line.
This essay argues for the other side. Forecasting India's overnight rate accurately is interesting; using what we learned about that rate to change how it is steered is the actual point. We were doing this for a Data Science & Management course, and the management half of that brief means asking a simple question: if the findings are real, who acts on them, and how?
2. How the system works today
The Reserve Bank of India sets a policy stance and publishes a repo rate. Every business day, scheduled banks meet their reserve obligations by lending each other money overnight; the rate at which those trades clear, weighted by volume, is the WACMR. The two numbers are tied together by arbitrage — banks will not pay more than the corridor ceiling (MSF) or accept less than the floor (reverse repo) unless their liquidity needs force them to.
The gap between WACMR and the Repo Rate is therefore a clean, daily verdict on whether RBI's stance is actually transmitting into the banking system. Today, almost every dashboard, treasury report, and news ticker we surveyed publishes WACMR as a level. The spread — the only quantity that distinguishes a stance that is biting from one that is announced — is buried.
3. What our findings expose
Three findings translate directly into policy facts. First, the rate corridor dominates:roughly 90% of our model's predictive signal comes from last week's WACMR, the Repo Rate, the MSF, and the spread itself. Equity and forex features do not appear in the top fifteen. Transmission is structural, not narrative.
Second, March 2020 was a regime break, and it outlasted the pandemic. K-Means did not know about COVID; it found the break in the feature geometry. The 2022–24 re-tightening did not return the system to its pre-2020 state.Third, the model's response function is asymmetric — predicted cuts transmit faster than predicted hikes of equal magnitude. This is a learned pattern from the data, not an assumption.
4. Three operational levers
Publish the WACMR–Repo spread as a headline statistic.Our SHAP analysis is unambiguous: the spread is a top-five predictor of next week's WACMR; the level alone hides regime-conditional behaviour. Surfacing the spread on the RBI's weekly statistical supplement, and including it in major financial news cards alongside Repo and CPI, would change what bank treasurers and policy analysts watch. The cost of implementation is roughly zero; the data already exists.
Adopt regime-aware briefings inside the policy desk. The post-2020 regime is not just a level shift — variances, spread distributions, and feature importances all changed. Internal MPC documents and RBI bulletin commentary should explicitly state which regime is currently active, and which historical comparisons are valid. A 2018 transmission analogue is not a useful guide to a 2024 decision.
Spend the communication budget asymmetrically. Our counterfactual response curve shows hikes transmitting more sluggishly than cuts. If the goal of MPC communication is to anchor expectations, the marginal speech, footnote, and press conference is worth more around tightening cycles than around easing ones. This is the opposite of how communication usually scales — easing cycles are politically harder to explain, so they get more airtime — and it is the kind of recommendation that only falls out of looking at the data.
Drag the Repo-rate slider, watch the predicted WACMR move, and inspect per-feature attribution.
5. Impact and tradeoffs
Three real tradeoffs sit behind these recommendations. Publishing the spread risks reflexivity — once a number is watched, it is also tradeable, and a positive spread can become self-reinforcing if liquidity desks treat it as a panic signal. Regime-aware language is harder to teach than a Phillips-curve story; the public communications team would have to argue every quarter why the active regime is what it is, and they would not always win. And our model is descriptive, not causal: the response function it learned is a fingerprint of the past ten years, not a guarantee of the next one.
The empirical basis for every claim in this essay — actual vs predicted, residuals, and feature importance.
6. What changes in practice
A bank treasurer watches the spread, not just the level. A Ministry of Finance policy desk asks which regime its assumptions are anchored to. The MPC communications team allocates more drafting effort to tightening announcements than easing ones. A market analyst writes the spread into the lede of every weekly column. None of these are revolutionary acts. Each is one reading habit, replaced.
Research-to-policy is a long pipeline. This is one brick in it.