Time series decomposition and analysis Using Python

Рет қаралды 25,327

Күн бұрын

#timeseries #decomposition #analysis
The decomposition of time series is a statistical task that deconstructs a time series into several components
Trend component - which reflects the long-term progression of the series - Trend can be positive or negative or both
Seasonal Component - includes cyclical component
Noise or residual - remainder of the time series after the other components have been removed
AR and MA model assumes time series to be stationary and real-world data - they are often
governed by a (deterministic) trend and they might have (deterministic)
cyclical or seasonal components

Пікірлер: 28

@vipulgaurav3813 4 жыл бұрын

Wonderful content! Following till the end gained many insights! Thanks

@RajeshGupta-gx3yz 3 жыл бұрын

hi Srivatsan, great video as always! Just want to know if there is a way to automate the process of identifying if a specific time series has seasonality (additive or multiplicative) and subsequently identify the period/season? Any help is much appreciated

@ruchigupta8378 3 жыл бұрын

this is really helpful, best series so far :)

@deeptigoyal4342 4 жыл бұрын

One of the best explanation. :)

@startupandgrowth 2 жыл бұрын

In the decomposition you have shown the seasonality can be either additive or multiplicative. What about the trend? Is it chosen same as seasonality? What if I want to decompose a timeseries with additive trend and multiplicative seasonality?

@sagarrawal8332 Жыл бұрын

How to get a value from graph such as 200 revenue/year and how to use linear regression to get value that is closely match with the observed value.

@vidyakurada4728 4 жыл бұрын

Thanks. I typically worked with multivariate ARMAX but in engineering problems where we largely prefer additive models. There simple mean deduction helps with de-trending and stationarity. I never saw time series in this level of non-linear multiplicative perspective. Currently I have been working on time series classification. It would help if we get to have a brief on that. Right now, I am approaching classification with regular ML models with time series features. Can you please suggest if it is alright?

@AIEngineeringLife 4 жыл бұрын

Most data are additive in nature. Even multiplicative can be made additive with log transform in many cases. Yes classification with TS features are very common when we do behavioral models or RFM or fraud/risk models and many more

@sharatchandra2045 3 жыл бұрын

Well explained

@pokepoke3 4 жыл бұрын

Thank you! From what I understand, seasonal_decompose uses moving average to determine the trend line. Do you know if there is anything wrong with using, say, a polynomial regression to determine the trend?

@AIEngineeringLife 4 жыл бұрын

Yes it used MA for trend. You can use polynomial if the trend is monotonic in nature over time. If there is seasonality as well as variance in trend not sure if PR will give a right fit. I have not tried if there is variant of PR that achieves it

@pokepoke3 4 жыл бұрын

@@AIEngineeringLife Thank you! How does MA determine the trend after the dataset? If we see a trendline going down, how does it know how much it will go down by later on (after the dataset)? Actually, what I want to know is... say we use ARIMA. How does ARIMA then factor in that the trend is going up or down?

@AIEngineeringLife 4 жыл бұрын

@@pokepoke3 MA models might not be able to capture trend reversal unless it happens. That is the disadvantage of of it. But if the trend reversals are seasonal then SARIMA can help and if trend reversals can be predicted with other external variables then multivariate analysis might help

@pokepoke3 4 жыл бұрын

AIEngineering Thank you very much. One last question. When we use something like SARIMA, I understand that I/d is for the order of difference. However, how does this apply for out-of-sample forecast? How does the model “un-difference”?

@AIEngineeringLife 4 жыл бұрын

@@pokepoke3 .. Basically undifference is the reverse transformation applied back to the last datapoint before out of sample that we know. So if you are applying First order Differencing - y′t = yt − yt−1 then you just add the predicted value back to last original value. It is coming in my upcoming videos as well

@sahibdeepsingh6969 Жыл бұрын

What's the y-axis in case of trend line and seasonal curve after decomposition?

@hitarthmukundraykanakia3137 3 жыл бұрын

Great video!! I had a question though. What do we mean by "seasonal fluctuations increase with the level of time series"? What is a "level of a time series"? How do we quantify fluctuations?

@pinkiki9972 3 жыл бұрын

Sir, i love your video series, best time series application explanation. I do have a question on how do I decompose the time series seasonality if my data set is a series of "life span" of tooling and it is not tie to datetime, but rather the life of the tool. The beginning of tool life vs 25%,50%,75% of the lifespan do demonstrate a seasonal pattern.