How a flood becomes a loan default
Four channels carry a climate shock from a field or a street to a bank's non-performing loans, and why measuring them takes more than a rainfall chart.
A bank does not see a flood. It sees a borrower who is late on a payment. Between those two events there is a chain of decisions, and the chain is where climate risk becomes credit risk. My BA thesis tried to measure that chain for Pakistan. This post sets out how I think about it.
Four ways a shock reaches a loan book
1. Income. A farmer whose crop is damaged has less to repay with. A shopkeeper whose street is under water has no sales for a week. For small borrowers, one bad month can turn into a missed instalment.
2. Collateral. Land, stock and buildings back many loans. If a shock lowers their value, or makes them harder to sell, the bank recovers less when a loan goes bad.
3. Concentration. Lending is not spread evenly. A district or sector that depends on one crop or one trade means one shock hits many borrowers at once, and their defaults arrive together.
4. Policy response. Governments and central banks often respond with relief, rescheduling or forbearance. That can delay defaults and then bring them back later, which makes timing hard to read in the data.
What a bank can and cannot see
The standard outcome is the non-performing loan (NPL) ratio, usually the share of loans that are 90 days or more overdue. It is a lagging measure. A shock in one month may show up as an NPL several months later, and the delay differs by sector.
The climate side has its own problem. Weather station records are sparse in mountain and remote districts. Satellite data helps because it covers every district every month: rainfall estimates, vegetation indices that track crop stress, and flood extent maps. That is what allows a district-month panel instead of a handful of national averages.
Why ARDL
When you mix climate series and financial series, some are stationary and some are not. An autoregressive distributed lag (ARDL) model can handle a mix of I(0) and I(1) variables, as long as none is I(2). Written in error-correction form it separates the short run from the long run:
$$\Delta y_t = \alpha + \sum_{i=1}^{p}\phi_i\,\Delta y_{t-i} + \sum_{j=0}^{q}\theta_j\,\Delta x_{t-j} + \lambda_1 y_{t-1} + \lambda_2 x_{t-1} + \varepsilon_t$$
Here $y_t$ is the credit outcome and $x_t$ is the climate shock. The $\theta_j$ terms describe the short-run response. The $\lambda$ terms, tested with the bounds approach of Pesaran, Shin and Smith, tell you whether a long-run relationship exists at all.
What I would check next
- Does the effect differ by sector, for example agriculture against trade?
- Does it grow with how concentrated a district's lending is?
- How long is the lag, and does relief spending change it?
These are the questions a credit officer or a regulator would also ask, which is why I think climate risk belongs in ordinary credit analysis and not in a separate report.
This is an explainer. It describes the method, not the results of any specific model.