How to associate daily rainfall probability with historic data Given a rainfall forecast of d days ahead and historic data collected over y years, what's a simple (but correct) way to associate the rainfall probability for each forecast day? 
I know this is a complex problem, but my purpose is to quantify the probability of a given rainfall volume 'happen' (ie, observed rainfall >= forecast for that day).
My first take was to use a simple frequency approach over the years:
target_day = 15
target_month = 1 
target_day_forecast = 5

data = select over years such that day = target_day AND month = target_month
n = number of days where forecast >= target_day_forecast
total = number of points in data

return n/total

On the other hand, i had trouble fitting a gamma distribution:
 require(MASS)
 fit = fitdistr(serie, "gamma")

 #Error in optim(x = c(0, 0, 0, 0, 0, 0, 0.3, 0, 0.1, 0, 0.8, 0, 0, 0, 0,  : 
 #initial value in 'vmmin' is not finite

If a sum a small quantity to serie, it works:
fit = fitdistr(serie + 0.1, "gamma")

My idea is to fit a gamma for each month, then i can extract probabilities easily. So, how can i handle the gamma fit problem above?
I also appreciate any ideas on this, 
Thanks!
 A: First, it bears mentioning that, yes, this is a complex topic, so 
Also, we should address the word "correct" here. Given that a particular method is valid, there are still some methods that are going to be more effective than others. It sounds like when you say "good," you really mean "valid." The "good" way to do this involves some fairly complex climate modeling, but there are still "valid" approaches we can take that, although the accuracy of the results will be limited, will still give us some usable results. Moreover, I'd suggest that there's a difference between the probability based on historic data and the "forecast." 
If you want a very simple forecasting model, you could just use the rain output (or the weather in general) from the day before. If you want a historical model to generate a probability, you could use the distribution of rainfall volumes for that date historically. These methods are naive, but they're valid and can produce reasonable results. 
If you want a model bordering on accuracy, you'll need to get pretty fancy. 
