Another great video Egor! Well done

@egorhowell

11 ай бұрын

Cheers David!

4:18 The seasonality and residuals look much multiplicative. box-cox says to transform skewed data into normal. One thing that confuses me is whether box-cox also turns multiplicative ts into an additive ts, because 5:44 after transformation, seasonality and residuals seem additive.

@egorhowell

10 ай бұрын

Yes you are right, the box-cox transform can also be used to turn the time series to be additive. I discuss this in another video in this course: kzread.info/dash/bejne/Zoqlmq2qkquxmJs.html

Thanks Egor.Its very understandable:) Helps a lot! Is there a way to reduce the effect of outliers on the transformation?

@egorhowell

10 ай бұрын

Thanks Anton! Hmm I am not too sure, my advice would be to remove them when calculating the box-cox lambda value. However, let me research and I will send some resources your way! :)

@antonschulte2538

9 ай бұрын

@@egorhowell Thanks a lot:)

@egorhowell

Ай бұрын

No problem

Another thing that confuses me is that box-cox says to transform skewed data into normally distributed. How can a time series show skewness? For cross-sectional data, the distribution of Y at a given x (draw a vertical line in the plot) is meaningful, because there can be multiple ys given an x, so we can see how frequent Y takes some value. While time series are essentially different, because a historical number has been observed and unique, I cannot figure out a distribution by a single certain value. Then how is it meaningful to apply box-cox to transform skewed data into normally *distributed*?

@egorhowell

10 ай бұрын

I understand your point, as for a normal regression problem we can see how the target variable is distributed relative to some feature. The goal is not necessary to show the skewness, because each data point belongs to a different distribution, so in reality it's quite difficult. However, I must say that this is reaching the edge of my current knowledge, so will be an interesting one to research about! Thanks for the question.

hey, are you supposed to apply the boxcox after applying log and diff or instead of log?

@egorhowell

6 ай бұрын

Hey! Normally you would apply the Box-Cox transform and then differencing transform. In reality, it depends what you are after. - diff(boxcox(x)) calculates relative changes - log(boxcox(x)) calculates the absolute differences Hope that makes sense!

Hi Egor, let say my end goal is to train, validate and test on monthly data from Jan 2023 to June 2024 and writing a function that takes user input of a month possibly next month(July 2024) to predict total monthly sales amount. if I apply box cox or any transformation the range of original value is going to be changed (Mostly going to be decreased). Do you think we must revert back the transformation when user gives input month for the prediction?

@egorhowell

21 күн бұрын

Yes, your prediction is going to be the box cox transformed output. So you need to do the inverse box cox to get back to the original time series units

Hi @Egor: Thank you for the videos. While going thru 3rd video, I saw the plotting function missing the following code. def plotting(, text=False, lam=None): if text: fig.add_annotation(x='1952-12-20', y=10, text=f'Lambda = {lam:.3f}', align='left', yanchor='bottom', showarrow=False, font=dict(size=20, color='black', family='Courier New, monospace'), bordercolor='black', borderwidth=2, bgcolor='white')

@egorhowell

5 ай бұрын

Hey, thanks for noticing this! I see this in the video, are you referring to the one GitHub?