I'm building an model on different loans' NPL rate. The problem is NPL rates are always affected by the market. When NPL rates move in trend, my model will fail the back-testing.

Assuming $x(t)$ is a random variable that distributed among $[-1, 1]$, with the mean $\mu = 0$ and a standard deviation $\sigma$.

When the sample size $n$ is big, the distribution of observed mean $\bar x$ will be ~ $N(0, \sigma^2/n)$. The back-testing confidence interval for $\bar x$ is $[- z_{\frac{\alpha}{2}} \frac{\sigma}{\sqrt{n}} , z_{\frac{\alpha}{2}} \frac{\sigma}{\sqrt{n}} ]$. The model always passes the backtest.

Now, troubles come when $x(t)$ has some trend. Let's say $x(t)$ becomes $x(t) = \sin (t + random(t) )$. Here $x(t)$ still follows the distribution, but when $x(t)$ moves near $+1$, the samples' mean will be around $+1$, increasing the sample size will not bring the $\bar x$ near to $0$, the model fails the backtesting.

Now my problem is, NPL's trend is hard to predict, the cycle sometimes are 6 months, sometimes are 2 years. My NPL model always fails the back-testing because of trend.

Any suggestion please?

  • $\begingroup$ You are basically asking people to design a strategy for you, hardly the purpose of this forum. If your strategy fails when the underlying driver is not mean-reverting then you need to work on a change in strategy approach if you want it to handle environments of high auto-correlation. $\endgroup$
    – Matt Wolf
    Jul 28, 2013 at 8:07
  • $\begingroup$ @MattWolf I would assume this is a general problem that many people has met? I'm not asking for an off-the-shelf solution, but looking for some experience to be shared. for example, once my model of saving account always fail due to some "elephant account", someone shared with me how to define and exclude outliers. $\endgroup$
    – athos
    Jul 28, 2013 at 8:46
  • $\begingroup$ You could research "regime changes" and the handling of those. $\endgroup$
    – Matt Wolf
    Jul 28, 2013 at 13:03

2 Answers 2


In my experience with forecasting, you could try a model of the form $$ X_ t = cycle_t + seasonality_t + residuum_t. $$ Sometimes it is hard to find the cycle but the seasonality could be doable if it has some natural structure (something happening in a certain month each year e.g.). Rob Hyndman explains all these things (and provides an R package) in his free online book. You could have a look at chapter 6. After having calculated the first 2 components you could model the distribution of the residuum.


you could calculate and subtract the trend.

x_(t) - x_(t - 1) = m (1 - exp(- h Dt)) + (exp(- h Dt) - 1) x_(t - 1) + e_t
error e_t normally distributed with

(s_e)^2 = [1 - exp(- 2 h)] (s^2)/2h

Because when Dt->0

x_t - x_(t - 1)  ~h(m-x_(t - 1))dt +  et 

you calculate the parameters h and m from the mean regression after that....?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.