Evaluating a CDS Pricing Model Through Hedging P&L

Summary

The document asks how to assess a single model that estimates credit default swap prices against historical data, beyond relying on root mean squared error. Its response recommends evaluating the hedge implied by the model: construct the associated hedge and measure how well it performs when the CDS portfolio is rebalanced daily. This shifts evaluation toward the financial task the model is meant to support, rather than judging only price prediction errors.

The answer also recommends pairing ordinary statistical goodness-of-fit measures with profit-and-loss-based measures, including the Sharpe ratio and a proposed replication-accuracy measure. It notes that, in some simple settings, P&L measures can be deterministic functions of familiar statistical fit measures. That qualification means the measures are not always independent evidence. The excerpt does not define the replication measure, prescribe a particular statistical metric, or provide empirical results, so applying the advice requires specifying the hedge, rebalancing assumptions, and evaluation period.

Key ideas

• Assess a CDS pricing model by testing the hedge it implies on a portfolio.
• Daily hedge rebalancing is proposed when daily observations are the available data frequency.
• Financial model evaluation can include both statistical fit and P&L-based measures.
• Sharpe ratio and replication accuracy are named as possible P&L-based measures.
• In simple cases, P&L measures may be determined by statistical goodness-of-fit measures.

Tags

# Single Model Accuracy Estimation


# Single Model Accuracy Estimation












I'm working on a model to estimate CDS prices, and want to backtest it against a historical timeseries. What are some error/goodness of fit measures that I can use for this purpose outside of RMSE?

I'm generally unfamiliar with this kind of metric, and in my research I've only found comparison measures, that yield a 'best' model relative to others, such as AIC and BIC. In this case I only have the one model, and want to produce some standalone measure of 'accuracy'.

## Answer by stans (score 1)

https://quant.stackexchange.com/a/41339

A standard approach here is to build a hedge implied by your model and evaluate its hedging performance when it comes to daily rebalancing of your CDS portfolio... I assume daily data is the highest frequency you've got.

You are doing finance, right? So in addition to regular statistical goodness-of-fit measures, you should always try PNL-based goodness-of-fit measures, whether it's Sharpe ratio or the "replication accuracy" measure I have proposed.

Having that said, in some simple situations PNL-based measures can be proven to be deterministic functions of the good old statistical goodness-of-fit measures.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.