Forecasting Extreme values I am attempting to forecast "peak" values. I have a weekly data set ~10 years. I have weather metrics as explanatory variables. I am trying to produce a model that can accurately predict the peaks of the weekly data using the weather metrics. The peaks are not entirely accurate, should I be using another approach? I am currently using MLR. Pred vs actual shows good explanatory ability until you reach more extreme observations. 
Thanks
 A: I do not have the power to comment, so I answer.
If the model has a "radius of convergence" outside of which its predictive ability is essentially meaningless then I would infer that there is a second model that is not informed by your current data, that governs in that extreme range.
Questions whose answers might be useful include:


*

*what is the exact edge of the "convergence range"?  

*Can you find explanatory variables that indicate smaller but specific and accurate variations in the edge of the range?

*Is the sampling once per week too sparse?  Is the under-sampling causing the misinformation.  


You can go to places like weather underground and get hour-by-hour weather for particular locations.  You might compare the predictive power of weekly data with the predictive power of daily data.  If the daily data gives substantially better results then at least part of your problem is undersampling.  You might also consider adjacent geographies.  When a storm hits down, it first had to go through another place first.  
