Methods

Every measure we construct, rather than copy from an agency, is listed here and tagged Original on the site. Each entry covers: what it is, why it matters, inputs (series ids), method, limitations, and validation.

Live

Core CPI and the "fuel" index

• What: CPI excluding food and fuel, using RBI's definition so it matches what the MPC looks at. "Fuel" = electricity, gas and other fuels (group 04.5) + fuels and lubricants for personal transport equipment (class 07.2.2: petrol, diesel, CNG). Source of the definition: RBI Bulletin Table 19(c), footnote #, which points to Box IV.2 of the April 2026 Monetary Policy Report.
• Series: cpi_core_index, cpi_fuel_index (pipelines/sources/derived/core_cpi.py).
• Weights (registry/weights/cpi2024_core.yaml): food and beverages 36.75 (MoSPI), core 53.04 (RBI), fuel 10.21 = the remainder. MoSPI does not publish the split of fuel between its two parts in machine-readable form, so scripts/estimate_core_cpi_weights.py recovers it. CPI 2024 is fixed-weight, so each parent index is an exact weighted average of its children, and constrained least squares over 21 months gives the shares (max residual 0.007 index points). The estimates are 5.372 + 4.848 = 10.220, scaled to sum to 10.21.
• Formula: core = (100·general − 36.75·food − 5.367·elec_gas_fuels − 4.843·vehicle_fuels) / 53.04.
• Validation (2026-10-28): vs RBI's published core index for Jun-25, Jul-25, Jun-26 and Jul-26, within 0.01 index points; FY26 average 104.31 = 104.31; y/y Jun-26 3.90 = 3.90, Jul-26 3.87 vs 3.86. Pinned in tests/sources/test_core_cpi.py.
• Limitations: starts January 2025 (MoSPI's back series exists only for the general index), so y/y starts January 2026. Re-run the weight script if MoSPI revises weights. Weights are fixed until the next base revision.
• Contributions: the inflation page shows the percentage-point contributions of food, fuel and core: w·(I_t − I_t−12)/General_t−12. They sum to the published headline rate (checked: Jun–Sep 2026 within 0.01 pp).

Brent crude in rupees

• What: Brent (US$ per barrel, front-month future, Yahoo Finance) × FBIL's USD/INR reference rate. Series brent_inr (pipelines/sources/derived/brent_inr.py).
• Method: each Brent close is converted at the reference rate of the same date, or the latest earlier one within 5 days. Days with no rate that recent are left out, not filled with a stale rate.

Growth rates "as published" across base years

• What: the headline growth series for CPI (cpi_headline_inflation), IIP (iip_general_growth) and GDP/GVA (gdp_real_growth, gva_*_growth) show each period's rate as the agency published it.
• Method: where the new base (CPI 2024, IIP 2022-23, GDP 2022-23) has a rate, it is used. Earlier periods use the old base's published rate. Nothing is recomputed or spliced, and the note column on every row says which base it comes from.
• Limitations: rates from two bases differ in basket and weights, so a long chart joins two definitions. Charts say so in their notes.

Year-on-year changes and other transforms

• What: growth rates computed from published index levels (e.g. CPI components, IIP categories, trade) are labelled "computed from the indices" on the site.
• Method: % change against the same calendar period a year earlier (matched on the date, not the row position). Because published indices are rounded, a computed rate can differ from the published one by about 0.01–0.1 points. Where an agency publishes the rate itself, we show that and label it "as published".
• 7-day averages (7dma, 7dma_yoy): for noisy daily series such as electricity demand, a 7-day calendar-window mean (at least 5 observations), and its % change on the same date a year earlier. Moving festivals and weather still distort the year-on-year comparison.

GST growth on a consistent definition

• What: gst_gross_growth, gst_net_growth.
• Method: each month's growth compares that month with the year-earlier column in the same sheet of the GST portal's workbook, so both sides use one definition (later sheets exclude the compensation cess). Levels are taken only from each month's own sheet.

Planned

Linked series across base years (Phase 2)

• What: long series for GDP, CPI and IIP across the 2026 rebasing.
• Method (default): ratio splice at the overlap period (back-cast the old series by the ratio of new to old in the overlap year or month). Document the overlap used per series. Do not splice growth rates.
• Limitations: composition changes between bases. Show the overlap comparison on the Data & Methods page.

Seasonally adjusted momentum (Phase 2)

• What: 3-month saar for CPI and IIP.
• Method: X-13ARIMA-SEATS (statsmodels wrapper) with Indian festival adjustments (Diwali moving holiday) where significant. Parameters stored in the registry.

Activity and inflation surprise composites (Phase 4)

• What: z-scores of each indicator's latest reading vs its own history (rolling 10y), averaged into an activity composite and an inflation composite. India's version of econvitals' "relative to past performance".

MPC Speak (live)

• What: a hawk–dove score for each MPC member's statement in each set of RBI minutes, since October 2016, plus the committee average (mpc_member_tone, mpc_committee_tone) and each member's vote on the rate resolution (mpc_member_dissent, mpc_dissent_count).
• Method: each statement is split out of the minutes by matching the "Statement by" headings to the names in the vote table. We count phrases from a fixed lexicon (registry/mpc/lexicon.yaml, version 1; regular expressions, case-insensitive, whole words). Score = (hawkish − dovish) / (hawkish + dovish + 2), so above zero leans hawkish and below zero leans dovish. The +2 keeps a statement with one or two hits near zero. No language model is used. That replaces the earlier plan for an LLM-scored rubric, for cost and so that anyone can reproduce a score from the text.
• How it was built: lexicon v1 was written before any scores were seen. The +2 shrinkage was added after looking at the hit counts of one meeting (August 2026, one to five phrases per statement), before the full history was scored.
• Checks (28 Oct 2026, 62 meetings, using only the minutes themselves): the decision comes from each rate resolution's wording ("to reduce / keep / increase").
  • Committee average by decision: −0.55 at the 13 meetings that cut, −0.10 at the 40 holds and +0.39 at the 9 hikes.
  • Correlation with decisions: 0.72 with the direction decided at the same meeting and 0.58 with the next meeting's.
  • Dissenters: among dissents whose statement says they wanted a cut (6 cases, by a simple "I vote to reduce/cut" match), the dissenter scored on average 0.30 below the median of the other members.
  • Caveat on the checks: a statement is written to justify the vote, so these checks show the score reads statements sensibly, not that it predicts anything.
• Limits: no negation handling ("no longer see upside risks to inflation" counts as hawkish). Members differ in length and style, so compare a member with themselves over time rather than members with each other. The vote table does not record which way a dissenter wanted the rate to move.
• Changing the lexicon: edit the yaml, bump version, add a changelog line, then run MPC_RESCAN=1 make update SOURCE=rbi_mpc_minutes (about 3 minutes) and make update SOURCE=derived_mpc_committee. Every changed score is logged in data/revisions/. Rerun the checks above and update them here.

RBI policy rule (Phase 4)

• To be specified.