Household Financial Fragility in India: A Composite Index and a Bayesian Early-Warning LSTM

Householdfinancialfragility—theinabilitytoabsorbmod- est income or expense shocks — is a central determinant of welfare in low- and middle-income economies, yet survey-based measures of it are scarce for Indian households. Using the sec- ond wave of the India Human Development Survey (IHDS-II, 2011–12, 𝑁=42,152 households),weconstructasix-component Financial Fragility Index (FFI) combining debt burden, con- sumption stress, asset deficit, employment concentration, depen- dency pressure, and a distress-borrowing flag; components are z-standardised, equally weighted, and quartile-binned into four fragility states. Three classifiers (logistic regression, random forest, isotonic-calibrated LightGBM) predict the binary Frag- ile/Distressed label from a disjoint non-financial covariate set, with a held-out ROC-AUC of 0.751 / 0.754 / 0.760 respectively and a leave-states-out generalisation AUC of 0.736±0.015. We then extend the cross-sectional baseline into a Bayesian tempo- ral early-warning system (Sect. 4): a macro-anchored monthly trajectory simulator with Poisson-distributed medical, disaster, and crop-failure shocks builds 24-month per-household panels; a two-layer LSTM with Monte-Carlo Dropout forecasts whether the household will cross the high-fragility threshold in the next three months; and a credible-interval decision rule converts the posterior into a Low / Medium / High triage tier. The LSTM attains a 3-month-ahead ROC-AUC of 0.982 — beating naive carry-forward, a per-household Kalman filter, and an NGBoost benchmark — and the resulting High tier captures 89% of all stress eventsat a 97% per-tierpurity.