Bridge ยท LAMF Cross-Sell & Unit Economics Engine

Borrow instead of break: the live findings

This page is generated from the actual pipeline output in Navneet-Scaler/bridge, hosted on GitHub Pages and rebuilt as the analysis changes. Every number below traces to a committed test or a query against the live database.

CI status version 112 tests passing python 3.12 license

The question

If BlinkMoney pushes LAMF harder, who should be targeted first, and what does it do to revenue?


Headline numbers

Median portfolio in this cohort
Rs 862
vs a Rs 25,000 industry minimum ticket
Propensity model, top predictor
5.06×
active_day_ratio, per 1 SD, vs 1.07× for portfolio value
Nudge effect on SIP breakage
−11.8pp
p = 0.0029, 95% CI −19.3 to −4.0

1. This cohort cannot support LAMF at real economics

A Rs 21/day micro-SIP book, one year in, has a median portfolio of Rs 862 and a maximum of Rs 11,841. The industry's typical minimum ticket is Rs 25,000. Three independent parts of the pipeline converge on this same wall: the simulation could not seed an eligible population at that threshold, the unit economics calculator derives a breakeven ticket of Rs 22,857 from cost structure alone (within 9% of the industry figure, by a completely different route), and a live query against the collateral table shows coverage falling to exactly zero at that price, not tapering toward it. Every revenue figure this project produces is an upper bound on a book that could not be written today.

2. Consistency predicts candidacy five times better than portfolio size

The propensity model (logistic regression, held-out ROC-AUC 0.870) found that active_day_ratio, sourced directly from Cadence's streak output, carries an odds ratio of 5.06× per standard deviation, against 1.07× for raw portfolio value. Investing reliably every day is a far stronger signal of LAMF candidacy than how much money is sitting in the account.

Odds ratio forest plot for the LAMF propensity model
Odds ratio per 1 standard deviation, held-out test set. Red bars raise candidacy odds, blue bars lower it.
FeatureOdds ratioDirection
active_day_ratio5.06×increases
streak_breaks2.37×increases
city_tier_tier_31.30×increases
portfolio_value1.07×increases
tenure_days0.60×decreases
kyc_incomplete0.43×decreases

3. The borrow nudge works, independent of the collateral problem

Among users who actually cleared the eligibility screen, showing "borrow instead of withdraw" cut SIP breakage from 45.2% to 33.4%, a change of −11.8 percentage points (p = 0.0029). The two arms were balanced on every pre-treatment covariate checked before the outcome was examined (p = 0.88 on eligibility, p = 0.42 on the size of the cash need), which is what licenses reading this as causal rather than as selection.

SIP breakage rate by treatment arm, intention-to-treat and among eligible users
Breakage rate by arm. Both the intention-to-treat estimate and the eligible-only estimate are shown, since quoting only the latter would overstate the intervention.

How the pipeline gets here

Cadence: streaks & transactions → Simulate portfolios & liquidity events → Leakage-safe features → Propensity model → Unit economics → Nudge validation → Weekly report & dashboard

Full architecture, the entity relationship diagram, and setup instructions are in the repository README. Bridge extends Cadence's shared Postgres database rather than standing up its own; it cannot run standalone, and the README says so directly rather than leaving that discovery to a reviewer.


Go deeper