Data

What is committed here

The files in raw/ are synthetic, produced by scripts/generate_data.py. They are calibrated to published characteristics of Gujarat monsoon rainfall but they are not observations. Nothing in this repository should be read as an analysis of real Ahmedabad weather.

File	Rows	Description
rainfall_ahmedabad_daily.csv	4,880	Daily rainfall, kharif window, 1985 to 2024
yield_ahmedabad_annual.csv	40	Annual district yield, kg per hectare

Calibration targets

Property	Value
Mean seasonal total	~690 mm
Coefficient of variation	~42%
Distribution shape	right-skewed, gamma-like
Trend	mild drying, roughly -1.9 mm per year
Dry-day probability	72%

Yield is generated as a saturating function of rainfall plus independent shocks. Those shocks represent pest outbreaks, heat stress, and harvest-time damage, none of which a rainfall index can observe. They are the reason the basis-risk analysis produces a non-trivial answer rather than a tautology.

---

Swapping in real data

The engine reads two CSVs with fixed schemas. Any source producing those schemas will work without code changes.

Schema

rainfall_ahmedabad_daily.csv

date,rainfall_mm
1985-06-01,0.0
1985-06-02,12.4

yield_ahmedabad_annual.csv

year,yield_kg_per_ha
1985,1642.0

Option 1: IMD gridded rainfall (recommended for India)

The India Meteorological Department publishes 0.25 degree gridded daily rainfall from 1901. The imdlib package wraps the download.

import imdlib as imd

data = imd.get_data("rain", 1985, 2024, fn_format="yearwise")
ds = data.get_xarray()

# Ahmedabad district centroid, approximately
lat, lon = 23.02, 72.57
series = ds.sel(lat=lat, lon=lon, method="nearest").to_dataframe().reset_index()

series = series.rename(columns={"time": "date", "rain": "rainfall_mm"})
series[["date", "rainfall_mm"]].to_csv("data/raw/rainfall_ahmedabad_daily.csv", index=False)

For a real product the grid cell should be replaced by the actual reference weather station named in the policy wording, since the contract pays on that station and no other. Grid-to-station mismatch is itself a source of basis risk.

Option 2: NASA POWER (global, no registration)

import pandas as pd, requests

url = (
    "https://power.larc.nasa.gov/api/temporal/daily/point"
    "?parameters=PRECTOTCORR&community=AG"
    "&longitude=72.57&latitude=23.02"
    "&start=19850101&end=20241231&format=JSON"
)
payload = requests.get(url).json()
series = payload["properties"]["parameter"]["PRECTOTCORR"]

df = pd.DataFrame({
    "date": pd.to_datetime(list(series.keys())),
    "rainfall_mm": list(series.values()),
})
df = df[df.rainfall_mm >= 0]
df.to_csv("data/raw/rainfall_ahmedabad_daily.csv", index=False)

Option 3: District yield data

Ministry of Agriculture and Farmers Welfare district-level crop statistics are published through data.gov.in and the Directorate of Economics and Statistics. Select a single crop, since aggregating across crops with different water requirements blurs the signal the basis-risk analysis is trying to detect.

For Ahmedabad kharif, cotton or groundnut are the sensible choices.

---

A note on record length

Forty seasons is thin for tail estimation. The 99th percentile of the fitted distribution is being extrapolated well beyond the observed minimum, and any pricing conclusion about the extreme tail should be treated as indicative rather than firm. Longer records, or pooling across climatically similar districts, would tighten it materially.