How to improve the accuracy of prediction based on past performance?

(post from old site)

One way is, instead of using the expected value, to calculate the confidence interval of the expected value and use the lower end of the value.

Another way is to find a period (period A)in the past when the conditions and the value similar to present, then use the period A+1 data to estimate the next period.

As usual, back testing with out of sample data.