Plotting actuals vs predicted by feature

In lesson 2 for Intro to ML, Jeremy asks to get an intuition of what good/bad prediction means. I tried to plot my actuals vs. predicted value for each feature on a Kaggle dataset.

The code used for the same is -

plt.figure(figsize = (15,10))
colslist = ['X314', 'X315', 'X8', 'X5'] # top most imp features
for i in range(1,5):
    plt.scatter(X_valid[colslist[i-1]], y_valid, s = 50, c='orange')
    plt.scatter(X_valid[colslist[i-1]],m.predict(X_valid),c='b', s=50, edgecolors='r') 

@jeremy I was wondering if the following insights can be drawn from the graphs that I see -

  1. The graph shows the distribution of y against a specific feature (in orange) and the red dots gives the distribution of our predictions for the same feature.
  2. From all the 4 graphs, it is evident that the value of y between 80-90 has not been captured completely in our prediction algorithm (m).
  3. The random forest algo doesn’t make predictions greater than 120 and considers them as outliers?