Computing confidence interval for average from individual predictions

In the attached image, I am plotting actual vs predicted along with confidence intervals for each of the predicted points. The black star represents the average of all the predicted points. What formula can I use to combine the confidence intervals of all the predicted data points to get a confidence interval for the average? Also, please assume that the average could be a weighted average.

I am using python, numpy.

• I'm guessing you want the "confidence interval of the mean" but it depends what question you are answering: you have expected vs predicted not expected vs actual or predicted vs actual, so I'm not sure what process is being modeled. Dec 30, 2016 at 18:08
• thanks @ctwardy, I should have been clearer, Where I say 'Expected', it means 'Actual'. Dec 30, 2016 at 23:10

OK, we have two dozen predictions-with-uncertainty, and actual point values. I'm still not sure we know enough to answer, so I'm going to just outline some points and hope someone can improve upon this. A quick search found three similar StackExchange questions:

• Combining 13 independent regressions seems the most similar -- Note comment warning this might be meaningless. The best-answer suggests a sampling approach because the result has no closed form. Using a Normal approximation would have a closed form, but I think your plot suggests the errors are not Normal, or your 95% CI is much too narrow.
• Combining 2 -- See esp. the reply discussing meta-analysis and whether you have a fixed effects or a random effects model. This may be what you are doing, in which case re-formulating the question for that community may be best.
• Delta Method -- Seems to me the same question, but takes a direct maximization approach, assuming approximations hold.

A couple of observations suggesting any simple average is not going to give you what you want, and you need to be very clear what is being combined, and what question the result is going to answer:

• If these are 95% Confidence Intervals, then your error model is much too conservative. Nearly all of them should include the true value (gray dashed diagonal) but only about 5 do. Therefore any parametric combination will be at least as unreliable. But suppose you have a black box with uncalibrated estimates, and you just want to know the "average" from that black box, however uncalibrated.

• If these are two-dozen unrelated estimates from a crowdsourcing platform, the "average" is not meaningful. On the other hand, if they are performance measures of the crowd, then the average is something like the average performance of the crowd across all question types. In that case:

• 1st approximation would just average the two-dozen points and use the standard deviation of that. (Each point can be weighted.)
• 2nd approximation would treat each estimate and CI as Normal, and use one of the formulas for combining Normal CIs. Note however that it seems error bars are wider towards the edges. That may require special handling.
• 3rd approximation: sample say 100 forecasts from the two-dozen distributions and then combine those ~2400 samples. (Weighting can be done by varying sample size for each forecast.) Works even if the distributions are not assumed to be Normal.
• If instead you have two-dozen forecasts from a single statistical model on different input data, and you want to know the average forecast across all input combinations, don't combine them. Just run the model with no factors specified, and use that result.

Hopefully a bona fide statistician can supply a more definitive answer.

You could also try a bootstrapping method in which you make the fit to only a subset of data points. This subset is created by randomly picking some of the data points. By repeating this many times you can take the uncertainty from the distribution of the fit results.