Hire me

Work/PRJ-07/Finance · India

UPI won. So why is India holding more cash than ever?

Because cash was never just for paying. UPI's yearly value grew from 37% to 91% of GDP in four years, yet cash in circulation is still about 12% of GDP, the same as in 2014-15, before UPI existed. UPI replaced cash for spending, not for saving.

Data
Reserve Bank of India statistics (4 tables), World Bank GDP
Period
2010-11 to 2025-26
Tools
SQL (DuckDB), Python (statsmodels)
Code
github.com/BoredMongoose/upi-cash-paradox ↗
Business question

UPI won. Should banks plan for less cash?

Recommendation

Plan cash logistics (ATMs, currency chests, note printing) for steady growth, not decline: cash is forecast to reach ₹48 lakh crore by May 2027.

Estimated impact1.1%

Average error of the cash-demand forecast in a 25-period backtest (forecast: ₹48.1 lakh crore by May 2027, 80% range ₹46.2–50.0).

01Results

UPI exploded. Cash didn't shrink.

UPI exploded. Cash didn't shrink. Both lines are shares of GDP. UPI's yearly value went from 37% to 91% of GDP, while cash held stayed near 12%.
Fig. 1UPI exploded. Cash didn't shrink. Both lines are shares of GDP. UPI's yearly value went from 37% to 91% of GDP, while cash held stayed near 12%.
Cash keeps returning to about 12% of GDP. Two shocks pushed it away: demonetisation (down to 8.9%) and COVID, when people hoarded cash while GDP fell (up to 14.8%). Both times it came back.
Fig. 2Cash keeps returning to about 12% of GDP. Two shocks pushed it away: demonetisation (down to 8.9%) and COVID, when people hoarded cash while GDP fell (up to 14.8%). Both times it came back.
UPI now handles 86% of non-cash transactions. Up from 63% four years earlier. Most of that share came from other digital channels, not from cash.
Fig. 3UPI now handles 86% of non-cash transactions. Up from 63% four years earlier. Most of that share came from other digital channels, not from cash.
UPI took over the small, everyday payments. The average UPI payment fell from ₹1,831 to ₹1,301, while debit-card payments fell 67% and their average size nearly doubled.
Fig. 4UPI took over the small, everyday payments. The average UPI payment fell from ₹1,831 to ₹1,301, while debit-card payments fell 67% and their average size nearly doubled.
Cash is forecast to reach about ₹48 lakh crore by May 2027. Holt-Winters forecast with an 80% range of ₹46–50 lakh crore. In a backtest it was off by 1.1% on average, slightly on the low side.
Fig. 5Cash is forecast to reach about ₹48 lakh crore by May 2027. Holt-Winters forecast with an 80% range of ₹46–50 lakh crore. In a backtest it was off by 1.1% on average, slightly on the low side.

02How I did it

From raw data to a decision.

  1. Read four messy RBI Excel tables into clean tables with Python.
  2. Built the analysis in SQL: staging tables, a consistent GDP series, then summary tables for each question. India changed its GDP base year in 2022-23, so older GDP values were rescaled using the year published on both bases. Without that, the cash-to-GDP ratio would show a fake 3% jump.
  3. Added 15 automated data-quality checks. For example, RBI's monthly and annual tables must agree for every March, and the individual payment systems must add up to RBI's reported totals.
  4. Compared three forecasting methods on 25 past periods before choosing Holt-Winters.

03What could make this wrong

The limits, said out loud.

  • 01Cash in circulation is a stock (measured at the end of March); UPI value is a flow (added up over the year). The chart compares their trends, not their sizes.
  • 02There's no direct measure of how often people use cash, such as ATM withdrawals. "UPI took the small payments" is inferred from ticket sizes and card volumes.
  • 03Official UPI figures in this source start in 2021-22.

04What I learned

  • 01Ratios can lie when a definition changes: India's GDP base-year change would have created a fake 3-point jump in cash-to-GDP.
  • 02Stocks and flows can be compared on trend, not on size.
  • 03Checking two official tables against each other catches errors neither shows alone.

05What I'd do next

  • 01Add ATM withdrawal data, to measure how much cash is actually used, not just held.
  • 02Break cash demand down by region and note denomination.
  • 03Re-run the forecast as each month's RBI data arrives and track its errors.
See the code on GitHub ↗All projects