2
$\begingroup$

I have dataset with product prices observed during period of time. E.g.

day_number | price
1 | 10
3 | 11
6 | 8
12 | 9
15 | 12
20 | ??

Price is recorded in data only on some days - it's missing for most of the days. If we assume that daily price changes are normally distributed with mean of zero and some standard deviation $X\sim N(0,\sigma_1^2)$, how can we calculate (based on data) 95% prediction interval for price on day 20?
I guess I need to first find most likely parameters for $X$ and then use formula for distribution of sum of random variables and get the prediction interval from there?

$\endgroup$
2
  • 1
    $\begingroup$ +1 Interesting question--and your proposed approach is good (although sensitive to the assumed normality of price changes, which will break down at low prices:you will need to check that assumption after estimating $\sigma_1$). Please note that your question asks for a prediction interval, not a confidence interval. But why did you apply the bayesian tag? And the sampling tag? If anything, time-series would be more appropriate, as well as missing-data. $\endgroup$
    – whuber
    Commented Aug 15, 2016 at 17:27
  • $\begingroup$ Thanks, I have edited the post according to your advice. I have initially set those tags based on my intuition about the problem, but I lack formal statistical knowledge so I guess those make no sense. $\endgroup$
    – hakaa
    Commented Aug 15, 2016 at 18:50

1 Answer 1

3
$\begingroup$

First, scale all the differences to obtain iid observations to estimate $\sigma_1^2$. For instance,

$Price(6) - Price(3) \sim N(0,3\sigma_1^2) => (Price(6) - Price(3)) / \sqrt{3} \sim N(0,\sigma_1^2)$

If we denote scaled observations as $y_i$, the estimate of $\sigma_1^2$ is:

$ \sigma_1^2 = \frac{1}{N}\sum\limits_{i=1}^N y_i^2$

Then you can get a CI using that

$Price(20) \sim N(12,(20-15) * \sigma_1^2)$

This approach ignores the uncertainty in $\sigma_1^2$ estimation. If you want to take that into account as well, you'll have to use simulation.

$\endgroup$
1
  • $\begingroup$ It's a good approach, but falls short at the end. It will be important to account for the uncertainty in the estimate of $\sigma_1^2$ because it's going to be large in examples like this. I believe there may be closed-form formulas for prediction limits. $\endgroup$
    – whuber
    Commented Aug 15, 2016 at 19:49

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.