VaR Backtest Starter

The scenario

You've joined the risk team. The desk runs a 99% one-day Value-at-Risk number every evening, and the regulator — and your boss — want to know it's honest. You're handed a history of daily portfolio returns and asked to estimate a rolling historical VaR and backtest it: how often did the actual loss blow through the VaR, and is that the right number of times?

Where this shows up

Task brief

# VaR Backtest Starter

**Role relevance:** Quant research take-home
**Estimated time:** 60 minutes
**Difficulty:** Intermediate
**Format:** Python (.py)

## What you are given
- returns.csv of daily portfolio returns
- var_backtest_starter.py with TODOs

## What you must deliver
1. Compute 99% 1-day historical VaR on a rolling window
2. Flag days where the loss exceeds VaR
3. Report the exception rate and a Kupiec test verdict

## Constraints & assumptions
Use a 250-day window. Do not use a parametric/normal VaR.

## Submission note
Complete the starter file, then compare your work against the mark scheme.

Your tasks

1. 01Compute a rolling one-day 99% historical VaR (as a positive loss) over a trailing window.
2. 02Flag exceptions where the realised loss exceeds the VaR known the day before (no lookahead).
3. 03Report the exception count and rate against the expected 1%.
4. 04Grade the model with the Kupiec proportion-of-failures test.

How you're assessed

The full points-based mark scheme is included with the pack.

What you'll learn

• ▸What a VaR number does and doesn't promise, and how it's estimated from history.
• ▸Why a backtest must use only information known before the day it judges.
• ▸How the Kupiec test decides whether the breach count is acceptable — and why clustering still matters.

Study alongside