4
$\begingroup$

This question is related to a rainfall prediction problem where I am trying to predict the amount of rainfall based on the meteorological features such as Max Temperature, Min Temperature, Dry Temp, Wet Temp, PET, Vapor Pressure, Relative Humidity, Mean Wind Speed, Wind direction, Brightness, Cloud and Pan Evaporation.

The challenges that I am facing are as follows, 1. Rainfall distribution is right-skewed, as it has more 0 values (no rainfall) than observations with rainfall. It is quite obvious as we have only 3-4 months in a year when rainfall occurs. If I go for log transformation of Rainfall (target variable), it doesn't make much of a difference. Because of large frequency of zero values, skewness still exist. At the same time, we can't ignore these values, we lose information there is no rainfall.enter image description here. 2. The scatter plot matrix between the features and target variable (rainfall) is displayed below. Correlations between the features vs target are low. Highest correlation has been observed < 34%enter image description here

How do I proceed in such case in order to build a predictive model? I tried many techniques such as log transformation, normalization as pre-processing, building RF model or neural network, none of the model is doing good, every model explains about 30-35% variability (r2 value). Any suggestions or opinions are welcome.

$\endgroup$
1
  • $\begingroup$ Welcome to Cross Validated and thanks for posting an interesting question. The quality of the answers you will get will likely improve if you further elaborate on the techniques that you have already tried out, i.e. if you show us what exactly you have done, what the results look like and what you think about them. $\endgroup$
    – Candamir
    Commented Nov 3, 2018 at 15:47

1 Answer 1

2
$\begingroup$

"It is quite obvious as we have only 3-4 months in a year when rainfall occurs"

You clearly need to split your regression problem between the rainy vs dry months. What about encoding a dummy variable for these months? Presumably a linear or tree model would do much better with this new feature.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.