Predicting with a GLM I wanted to check my understanding of predicting with a GLM:
A binomial/logistic regression model predicts the binomial parameter = p = P(success). To convert the probability into classes, we have to add a threshold or cutoff.
The same idea applies for a multinomial logistic regression model.
A poisson regression model predicts the poisson parameter = rate. To convert into counts, I use a threshold again?
A gamma model predicts the scale and rate parameters. I do not need a threshold because the response is continuous.
 A: Quoting from Wikipedia:

The GLM consists of three elements:

*

*An exponential family of probability distributions.

*A linear predictor $\eta=X \beta$

*A link function $g$ such that $E(Y|X)=\mu=g^{-1}(\eta)$

There is no threshold inherent in a GLM. Once you have the model, you can make predictions of $\mu$ (sometimes called the "mean function") for any set of covariates $X$. For a binomial model that could be translated into a probability of class membership. For a Poisson model, you are modeling counts directly.
Your application of a binomial GLM might then involve a threshold for making class predictions. Your application of a Poisson count model might involve translating counts into a rate per unit time, length, or area.  But those applications should be thought of as outside the GLM itself.
A: The commonality is that all these models predict conditional expectations. If your target class is coded 1 and your nontarget class is coded 0, then a predicted probability $\hat{p}$ of a new instance to belong to the target class is just the conditional expectation of the new instance's code. (Thresholding is iffy and loses a lot of information. Only do it if you know what you are doing.)
Your Poisson regression will also predict the conditional expectation. (As long as your prediction is on the response scale, not the linear scale.) You can feed this predicted expectation $\hat{\lambda}$ into a Poisson calculator to get predicted probabilities of each possible count. Note that this procedure is a shortcut that completely disregards the uncertainty in your estimate of $\hat{\lambda}$ - take a look here for a more stringent approach.
Note that there are other models that predict other functionals of the target variable's distribution, e.g., quantile regression, which aims at predicting a certain quantile.
A: It matters what you mean by prediction.  Unfortunately, this term can be somewhat ambiguous, especially since the linear combination of covariates in the regression model is often referred to as a linear predictor.
The typical purpose of a generalized linear model is to estimate the population mean and to perform inference on the mean.  This would be the proportion in a Bernoulli model and the mean in a Poisson or gamma model.
The word prediction is best reserved for when interest surrounds a future sampled observation.  Of course our best point prediction of a future observation is the estimated mean of the population.  For a gamma model one would report the sample mean as the point prediction for a future observation.  For a Bernoulli model one would report the value 0 or 1 that has the largest estimated proportion since an individual observation can only take on these discrete values.  For a Poisson model one could report the mean rounded to the nearest integer since the support of the Poisson distribution is the non-negative integers.  One could also use the floor or ceiling function on the mean to produce a point prediction.
One might also be interested in presenting the estimated percentiles of the population.  It is important that these be presented with tolerance intervals (confidence intervals for population percentiles).  Alternatively one might be interested in quantifying the uncertainty regarding the point prediction for a single future observation.  This would require the use of a prediction interval which is not the estimated percentiles.  Here is a related thread that discusses prediction intervals.
Addendum: Splitting the data into training and test is for the purposes of validating the out-of-sample prediction ability of a model.  My preferred approach is not to split the data into training and test sets.  Rather, I suggest to bootstrap (sample with replacement) $n$ observations from the data set as if it is the population, fit the model, and construct a point prediction or interval prediction for a particular prediction target (a single future $y$ [$m=1$ observation] or a future $\bar{y}$ based on $m$ observations).  Then bootstrap a sample of size $m$ and tally i) the discrepancy between the point prediction and the target, and ii) whether  the prediction interval covered the target .   Repeat this 10,000 times and plot the histogram for point prediction errors and calculate the coverage rate for prediction intervals.  This validates the performance of the model based on operating characteristics.
Sampling with replacement from your data set treats it as a much larger population.  It is likely the percentiles of your data set do not match the theoretical percentiles of the glm model you posit.  This means there is slight model misspecification so don't be surprised if the prediction intervals do not cover at the nominal level and if the histogram of prediction errors shows small bias (not centered at zero).  You can also perform this type of validation through simulation by randomly generating observations from the theoretical model that matches your glm, e.g. gamma or Poisson.  Here you should find the prediction intervals perform close to the nominal level and your point prediction is asymptotically unbiased for the target.
This type of approach can also be used to validate point and interval estimation of a population parameter.
