A fundamental idea in modern asset pricing is that risk aversion may be time varying and countercyclical (Campbell and Cochrane, 1999). Existing evidence is relatively scant and based mostly either on experimental/survey data or on stock market data directly. We aim to test this key assumption using a large panel of real-world transaction data. Risk aversion is imputed from Chinese auto insurance policies, including 10 million transactions from 2011 to 2017. We find that individuals' risk aversion is indeed time varying and countercyclical. The risk aversion measure behaves reasonably and also exhibits seasonal, "zodiac year," and calendar-event patterns.