Is this evidence my random forest model is over fit? I am investigating win probabilities in cricket and calculating the chances of winning a game depending at what stage the game is at. I created a logistic regression model and a random forest model. When I tested the models on a random game I got the following plot

They both follow the same broad trend but the random forest is a lot more variable. I am just trying to understand the causes of this and what I can do to improve it? 
 A: Without knowing the details of your model, it's very hard to give specific comments. However, in general a random forest can in principle have relationships of variables and outcomes, and interactions between predictors of arbitrary complexity. In contrast, a logistic regression with only main effects and continuous predictors used in a linear fashion (potentially after some transformation) will tend to be a less complex model that might be able (for better or worse) to react more strongly/in a more nuanced fashion to changes in predictors.
Additionally, you did not say whether your example is from the data you trained the model on or not. If it is, then a random forest is typically easier to overfit on the actual data than a logistic regression, but can also (when trained appropriately with well chosen hyperparameters) be somewhat better at predicting new data following the same distribution as the original data. If this is new data, that is a more meaningful evaluation. If you look at enough data that the models never saw, then you will find out whether the RF is overfit or not by seeing its performance on that new data. In the absence of new data, an evaluation using cross-validation may be the best thing you can manage.
A: As it's been pointed out already, to see whether a model suffers from overfitting one needs to choose a cross-validation scheme and compare training and test errors. For Random Forests you have two options.   


*

*Keep a subset of your data (usually 20%) aside, and train your model on the remaining set. Afterwards, you test the performance on your training and test data and compare these two numbers. This is applicable to any ML-model. 

*Alternatively, in Random Forests one can use Out Of the Bag (OOB) error estimation. As you might well know, Bootstraping is implemented in each tree in Random Forests, i.e. whenever a tree is built, a random subset of the training data is left aside (i.e. not considered in training). This is called an OOB subset. Using the OOB of each tree one can obtain the OOB error estimation. In modern ML API's like Sklearn, obtaining the OOB error estimation is rather straightforward. Assuming the numbers of trees is big enough, this can be shown to be equivalent to Leave-One-Out cross validation. (See An Introduction to Statistical Learning by Gareth James et. al.).
