ashwingopalsamy/india-repo-rate-dataset

India RBI Policy Repo Rate and Monetary Policy Decision History

An independent, reproducible dataset of the Reserve Bank of India (RBI) policy repo rate and its monetary-policy decision history. The canonical table is a decision ledger: each row is one dated rate observation, including unchanged policy decisions. Annual summaries, source metadata, contextual events, and repository-provided regime intervals are separate configurations derived from or accompanying that ledger.

<!-- BUILD-SUMMARY:START --> Current build: SnapshotV2 2026-08-14-v2, retrieved 2026-08-14T07:35:45.400Z; coverage 2000-06-05 to 2026-08-05; 107 canonical records, 27 annual rows, 9 sources, 8 contextual events, and 15 regime intervals. <!-- BUILD-SUMMARY:END -->

At a glance

• —decisions is the default configuration: one row per canonical policy-rate observation or identified policy decision.
• —effective_date orders the rate history; decision_date is null when a source does not identify a separate formal decision date.
• —policy_rate_pct is convenient numeric percent; policy_rate_bps is the exact integer basis-point representation.
• —change_bps, action, and is_rate_change describe the transition from the previous canonical record.
• —provenance_class and verification_status distinguish directly verified RBI records from legacy imported history.
• —record_text is deterministic factual text for search, RAG, and agent context; it is not LLM-written.

This dataset is designed for questions such as: What was India’s repo rate at the end of 2019? How much did RBI tighten during 2022? Which years had the most rate cuts? Which rows are formal MPC decisions rather than historical observations?

This is not an official RBI product. It is a derived research artifact created from repository-maintained source records; users should consult the linked RBI source document for authoritative policy information.

Why the decision ledger is canonical

The decisions configuration has one row per canonical SnapshotV2 record. It preserves hikes, cuts, and holds rather than reducing the history to one row per year. The annual configuration is calculated entirely from this ledger, so a year can contain multiple transitions without losing ordering, rate-before/rate-after semantics, stance, or provenance.

The Parquet files under data/ are the only files mapped into Hugging Face configurations. CSV and JSONL files under exports/ are deliberately not listed in the frontmatter, so they cannot be ingested by the Dataset Viewer as duplicate records.

Configurations

All configurations use a custom full split because this is a historical data ledger, not an ML train/test corpus. The exact Arrow/Parquet schema is available in `schema/`, and field-level semantics are in `schema/data-dictionary.json`.

Important field semantics

In decisions:

• —effective_date is the canonical ordering and rate-effective/observation date.
• —decision_date is populated only for the three RBI policy-resolution records in the current snapshot. Imported historical observations intentionally use null rather than an invented decision date.
• —policy_rate_pct is a convenient numeric percentage and policy_rate_bps is the exact integer representation. change_bps is signed: positive is a hike, negative is a cut, and zero is a hold.
• —previous_policy_rate_bps is the previous canonical ledger rate, not a previous row in an annual summary. The first observation has null previous rates and action initial.
• —canonical_key is stable and globally readable, for example IN:RBI:policy_repo_rate:2026-08-05. decision_id is a stable dataset-level identifier, while snapshot_decision_id preserves the source ledger ID.
• —record_text and decision_summary are deterministic factual text generated from structured fields. They are not LLM-written and do not repeat long RBI source prose.
• —regime_id joins to regimes using start-inclusive/end-exclusive intervals. Dates in a gap remain null. Regime labels are repository context, not official RBI classifications.

Source quality is explicit. official_primary identifies a direct RBI policy resolution or direct RBI rate-data source; official_secondary identifies RBI archive, minutes, or current-rate context; legacy_import identifies imported Reuters/Shriram historical material. verification_status makes the distinction visible as verified_primary_source, historical_legacy_source, official_secondary_source, or mixed_provenance. The current build contains 104 historical_rate_observation rows with legacy provenance and three policy_decision rows backed by RBI policy resolutions.

Annual calculation rules

The annual table includes every year from the earliest effective year through the latest snapshot year, including years with zero canonical ledger records. In the current build, 2021 is represented with decision_count = 0 and carries forward the rate in force from 2020.

• —start_policy_rate_pct is the latest rate effective strictly before January 1.
• —end_policy_rate_pct is the latest rate effective on or before December 31.
• —min_policy_rate_pct and max_policy_rate_pct consider known in-force rates, including a carry-in rate.
• —net_change_bps is end minus start and is null when either boundary is unknown. This is why 2000 has a null net change: the available history begins during that year.
• —gross_hikes_bps sums positive canonical transitions and gross_cuts_bps sums the absolute values of negative canonical transitions. Holds are retained in decision_count and hold_count; the initial observation is neither a hike nor a cut.
• —year_end_source_id supports the rate in force at year end, including a carry-in source for a zero-decision year. year_end_action and year_end_stance are null when no decision occurred in that year.

Load examples

The default configuration is the decision ledger:

from datasets import load_dataset

repo = "ashwingopalsamy/india-repo-rate-dataset"
decisions = load_dataset(repo, split="full")
annual = load_dataset(repo, "annual", split="full")
print(decisions.features)
print(decisions.filter(lambda row: row["action"] == "cut")[0])

For a local checkout or downloaded artifact, point data_files at the canonical Parquet file explicitly:

from datasets import load_dataset

annual = load_dataset(
    "parquet",
    data_files={"full": "hf-dataset/data/annual.parquet"},
    split="full",
)

Pandas and Polars preserve the typed Parquet columns:

import pandas as pd
import polars as pl

df = pd.read_parquet("hf-dataset/data/decisions.parquet")
cuts = df.loc[df["action"].eq("cut"), ["effective_date", "policy_rate_bps", "change_bps"]]

pl_df = pl.read_parquet("hf-dataset/data/annual.parquet")
print(pl_df.filter(pl.col("decision_count") == 0))

DuckDB can query the canonical file without loading a database:

SELECT year, net_change_bps, hike_count, cut_count, hold_count
FROM read_parquet('hf-dataset/data/annual.parquet')
ORDER BY year;

For RAG or an autonomous agent, retrieve decisions.record_text and retain canonical_key, effective_date, change_bps, stance, verification_status, and primary_source_url in the answer context. The sentence is deterministic, while the structured fields remain the source of truth for filtering and exact arithmetic. A source URL should be cited when a claim matters. Events are independent context and should not be described as causal explanations for decisions.

Source methodology and provenance

The build reads only src/data/snapshot.json, validates SnapshotV2, and does not use historical V1 snapshots or website runtime exports as additional inputs. The canonical rate transitions are checked with decimal arithmetic. Source IDs, source checksums, retrieval timestamps, and publication timestamps are preserved or normalized without replacing unknown values with sentinel strings.

The repository’s current history contains a Reuters historical series covering the older rate observations, with a Shriram Finance reference retained in the source table. Those rows remain useful for historical continuity but are visibly marked legacy_import and historical_legacy_source; they are not represented as equivalent to a row tied to a specific RBI policy resolution. The three latest policy-resolution records are marked official_primary and verified_primary_source. Source priority is deterministic and prefers a policy resolution over minutes, contextual pages, and legacy material when a row has multiple sources.

Reproduce and verify

From the repository root, install the pinned build/test dependencies and rebuild:

python3 -m pip install -r requirements-hf-dataset.txt
python3 scripts/build-hf-dataset.py
python3 -m unittest tests/data/test_hf_dataset.py
npm run test:hf-dataset

generated_at is intentionally the SnapshotV2 retrieval timestamp, not the wall clock, so rebuilding the same snapshot is byte-identical. The manifest records the snapshot ID/checksum, generator and schema versions, row counts, coverage, and output checksums. SHA256SUMS covers every artifact except itself.

The semantic dataset schema version is 1.0.0; a data refresh that adds a new policy decision does not by itself require a schema-version change. The source snapshot identity and freshness are separate fields.

Attribution & Usage

This is an independent, non-official educational reference compiled from publicly available records published by the Reserve Bank of India and other cited publishers. Source titles, marks, and publisher materials remain with their respective owners; this project does not claim ownership of third-party material. This project is not created by, affiliated with, authorised by, sponsored by, or endorsed by the Reserve Bank of India; no official relationship, approval, or representation should be inferred.

The information is provided for research and general information only, “as is” and without representation or warranty, express or implied, including as to accuracy, completeness, timeliness, availability, or fitness for a particular purpose, to the fullest extent permitted by applicable law. Verify every figure, interpretation, and update against the original publication before relying on it; this project should not be used as the sole basis for any decision.

Nothing here constitutes financial, investment, legal, tax, accounting, or other professional advice, an offer, recommendation, solicitation, or fiduciary relationship. Links and citations are provided for identification and verification only, not as a licence to reproduce protected third-party content. Nothing in this notice limits any right or remedy that cannot lawfully be excluded.

Reference material: RBI data dissemination material.

This notice is not a dataset license. No blanket redistribution license is asserted. The transformation code is separate from rights in RBI, Reuters, Shriram Finance, and other third-party source material. Review source-specific terms before redistributing or using the data commercially. The machine-readable rights summary is in `provenance/build-manifest.json` and `schema/data-dictionary.json`.

Limitations

The source records provide a strong machine-readable transition history but do not make every historical observation an identified RBI policy-resolution event. Older rows therefore have null stance and decision dates where the source does not provide them. Regime labels and contextual events are repository-maintained annotations, not official RBI classifications or causal evidence. Rate values and exact decision claims should be checked against the linked primary RBI source where one exists.

Suggested citation:

India RBI Policy Repo Rate and Monetary Policy Decision History, independent derived dataset from the repo-rate-visualizer SnapshotV2 ledger, schema version 1.0.0. Cite the snapshot ID, checksum, and the relevant source URL from provenance/build-manifest.json and the sources configuration.

Links

• —Dataset page
• —Source repository
• —RBI disclaimer and copyright policy
• —RBI data dissemination material
• —RBI historical data
• —Hugging Face manual dataset configuration
• —Hugging Face dataset file names and custom splits
• —Hugging Face Parquet Viewer processing
• —Hugging Face Croissant metadata