Suggest answer for theoretical question

[junjunbaby123]

Q: What if we want to build a model for predicting prices? Are prices distributed normally? Do we need to do any pre-processing for prices?


A: None of the models require independent or dependent variables follow normality assumption. The normality assumption is always on the error terms. We assume that after fitting the model, the error term are i.i.d. N(0, sigma2). 

Prediction of price could either be a time series forecast or cross-sectional forecast. In a time series forecast, we need to pre-process the price so that the time series is stationary. We can detect stationarity by using the ADF test. If it is not stationary, we use first-difference, second difference, ..., until the series is stationary. Sometimes we can use the log of price to resemble growth rate. Under the cross secitonal forecast, we can apply multivariate regression to predict price (such as house price forecast). Here we can also use the log of price, and we also need to engineer features and select features.

[Harshupadhyay221]

When building a model for predicting prices, it's important to consider several factors regarding the distribution of prices and any necessary preprocessing steps:
Distribution of Prices:
Preprocessing for Prices:
Feature Engineering:
Evaluation Metrics:

[Vishal123-tech]

A small clarification: prices themselves do not need to be normally distributed. In most predictive models, the normality assumption (if any) applies to the residuals/errors rather than the independent or dependent variables.

For cross-sectional prediction problems (e.g., house prices), common preprocessing steps include handling missing values, treating outliers, feature engineering, and sometimes applying a log transformation to the target when prices are highly skewed.

For time-series forecasting, preprocessing depends on the model. Classical models such as ARIMA typically require stationarity, which can be checked using tests like ADF and achieved through differencing and/or log transformations. However, many machine learning models do not require strict normality or stationarity assumptions.

Therefore, prices are often not normally distributed, and preprocessing should be driven by the modeling approach and data characteristics rather than a requirement that prices follow a normal distribution.