It's my first question, I am not so experienced in stats: feel free to point me to the right direction.

I am doing a regression to predict a price. I wanted to check the residuals (the difference between the target and the predicted value) for clues to improve the prediction. I note a linear pattern (see below), I am right in interpreting that I likely miss a linear predictor? What is the best way to fix such issue? Would it make sense to estimate the linear coefficient, and add it to the (linear) model?

The data is from there: https://www.kaggle.com/c/house-prices-advanced-regression-techniques/data
The data description is : https://ww2.amstat.org/publications/jse/v19n3/Decock/DataDocumentation.txt

The model is the https://www.tensorflow.org/api_docs/python/tf/estimator/DNNRegressor
for which I selected 10 predictors.

In parallel I also made a linear model: https://www.tensorflow.org/api_docs/python/tf/estimator/LinearRegressor which use the same predictors. The residuals also shows a linear pattern (together with the outlier), although the mean seems to be zero.

  • Residual for linear model:
    linear pattern in the residuals plot (linear)
  • Residual for non linear model:
    linear pattern in the residuals plot (DNN)
  • 1
    $\begingroup$ Welcome to the site. It looks like some thing is wrong here. Not only do you have linearity, but you have an outlier. Also, the mean residual should be 0 and it seems pretty clear your mean residual is positive. Can you describe your data and model in more detail? $\endgroup$ – Peter Flom Oct 4 '18 at 18:09
  • $\begingroup$ I added some mode description about the data and model. I feel less worried about the outliers as there are outliers in the data and so far I did nothing about it. I feel more concerned about the general trend. $\endgroup$ – call me Steve Oct 4 '18 at 19:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.