WHATwe'dBUILDnext

1. What we deliberately deferred

Every research project is mostly a list of things you did not do. Ours is no exception, and the omissions were not accidents. We chose a weekly grid because that is the native frequency of NDAP's RBI series. We stopped at two regimes because the silhouette curve told us to and we did not want to overfit a structural break that the data itself was indifferent about.

We hand-coded seventy-five policy events because manual scoring kept the noise floor low and the methodology defensible inside a one-semester course. We worked from a static July 2024 snapshot because shipping a live retraining pipeline was not in scope. Each of those choices was honest at the time. Each one bounds the answer the project can produce. This essay is the list of bounds, and what crossing them would cost — and reveal.

2. Five extensions in priority order

Daily-frequency pipeline. The weekly grid hides almost everything that liquidity desks actually care about. Daily NDAP fetches plus daily WACMR clearing prices would multiply the sample roughly five-fold and let us distinguish settlement-day stress from typical-day clearing. Most of the pipeline already runs on daily Yahoo Finance data; the lift is in re-doing the alignment layer and re-running every model with proper time-series cross-validation at the new frequency.

Hidden Markov regimes. K-Means on PCA gave us a clean two-state result, but the segmentation is hard — every week is in exactly one regime. A Hidden Markov Model with three or four states and soft assignments would let the 2022 tightening exist as a transient state instead of being absorbed into post-COVID accommodation. More usefully, an HMM emits a posterior probability of being in each state every week, which is exactly the early-warning signal a liquidity desk would want.

Causal identification. Our counterfactual simulator is a model-based counterfactual. It tells you how the XGBoost forecaster responds to a hypothetical Repo shock; it does not tell you what the Indian banking system would actually do. Closing that gap means high-frequency event studies around MPC announcements — looking at bond yields and interbank rates in five-minute windows around scheduled releases — and an instrumented-decision design that exploits the partial unpredictability of MPC voting.

LLM-assisted NLP at scale.Seventy-five hand-coded events is not a small dataset; it is a tiny one. The full RBI corpus — bulletins, MPC minutes, governor speeches, financial stability reports — contains thousands of policy-relevant text events over our window. A consistent LLM-based scoring pipeline, validated against our hand-coded gold standard, would replace the "news is a garnish" caveat with a real test of whether language carries marginal predictive content.

Live data and retraining. The dashboard is currently a snapshot. A weekly NDAP refresh job, drift detection on the feature distributions, an automated retraining cadence, and an alert when the regime-shift posterior crosses fifty percent would convert the project from research artefact to operational tool. The infrastructure work is straightforward; the discipline of running it in public for six months is the harder part.

3. Adjacent questions

Three questions sit one degree away from what we built and would each justify a project of their own. Predict the whole short-rate curve — MIBOR, T-bill yields, market repo, Commercial Paper — jointly rather than only the WACMR; the cross-correlations are where transmission lives. Compare across emerging markets — Brazil, Indonesia, Turkey, South Africa — to see whether the post-2020 break is an Indian phenomenon or a global one driven by central-bank balance-sheet expansion. Drop to bank-level granularity — anonymised lender and borrower distributions in the call market — to see who funds whom, in what stress, and whether the average transmission story hides important distributional asymmetries.

4. What this work would reveal

These extensions are not gold-plating; each one cashes out as a specific policy insight. A causally identified asymmetric-transmission claim would let the MPC defend its communication budget on quantitative grounds rather than intuition. A real-time HMM regime-shift posterior would give liquidity desks an early-warning signal that is currently delivered post-hoc, often months late. A measured communication elasticity — how many basis points a hawkish governor speech moves the spread, conditional on event type — would tell us whether forward guidance is doing the work the textbook claims it does.

5. Closing

A research project's value is not what it solves but what it makes possible. The data pipeline, the SQLite layer, the API, the dashboard, and the agent are all reusable. Future iterations should pick the question first — which regime is currently active, what would a 25 basis point cut do, why did the spread widen last week — and point the existing infrastructure at it. The next brick is the easiest one to add.