Aidemy 2020/10/29
Hello, it is Yope! I am a liberal arts student, but I was interested in the possibilities of AI, so I went to the AI-specialized school "Aidemy" to study. I would like to share the knowledge gained here with you, and I am summarizing it on Qiita. I am very happy that many people have read the previous summary article. Thank you! This is the fourth post in the next series analysis. Nice to meet you.
What to learn this time ・ Construction of SARIMA model
-The __SARIMA model __ is a model in which ARIMA (p, d, q), which converts data into a step difference series, can be converted to a time series with a __seasonal cycle. -The SARIMA model has a parameter called __SARIMA (sp, sd, sq, s) __.
sp,sd,sq -Sp, sd, sq are called __ "seasonal autocorrelation", "seasonal derivation", and "seasonal moving average" __, respectively, and are basically the same as p, d, q of ARIMA. ・ It was added that the data of sp, sd, and sq are influenced by the data of the past seasonal period. -The other parameter of SARIMA, __ "s" __, represents __seasonal cycle __. If it is a 12-month cycle, s = 12 should be set. -For sq, it indicates how many cycles ago the affected "past seasonal period" is.
-It is necessary to check appropriate values for the above parameters. ・ At this time, __Information Criterion __ is used. This time, I will use __ "BIC" __ among them. However, this time we will not deal with it in detail. -The lower the BIC value, the more appropriate the parameter value.
-__ Partial autocorrelation __ is __ without the influence of data between the two values that perform autocorrelation __. -For example, the partial autocorrelation between y1 and y7 is obtained by removing the influence of y2 to y6 between them. -By visualizing this partial autocorrelation, __ set the optimum value of the __parameter "s". -When visualized, the value of the partial autocorrelation becomes high in the __period part, so this should be the value of s.
-Visualization (graphing) of partial autocorrelation is performed as follows. __sm.graphics.tsa.plot_pacf (data) __
-Code (visualization of correlation of wine sales data)
・ Result
・ Construction procedure ① __Data reading __: pd.read_csv () ② __Data organization __: pd.date_range () ③ __Data visualization __: sm.graphics.tsa.plot_pacf () ④ __Data cycle (s) grasp __: From ④. ⑤ Parameter settings other than __s __: Grasp with BIC --New part from here-- ⑥ __ Model construction __: __ sm.tsa.statespace.SARIMAX (). Fit () __ ⑦ __Data prediction / visualization __: __ predict () __ / plt.show ()
・ ⑥ The arguments of SARIMAX () for model construction are as follows. __SARIMAX (data, order = (p, d, q), seasonal_order = (sp, sd, sq, s)) __
・ ⑦ The predict () of data prediction is as follows. Only the start of the forecast needs to be the time in the time series data. __ Model.predict ("prediction start", "prediction end") __
・ A series of code execution (⑤ is not done)
・ Result
-The __SARIMA model __ is a model that makes ARIMA (p, d, q), which converts data into a step difference series, compatible with time series that have a seasonal cycle. -The SARIMA model has parameters __ (sp, sd, sq, s) __. -For sp, sd, and sq, check the appropriate values using BIC, the __information criterion __. ・ S is a diagrammatic examination of __partial autocorrelation __. -Pass these parameters to __sm.tsa.statespace.SARIMAX (). Fit () __ to build a model and make predictions with __predict () __.
This time is over. Thank you for reading until the end.
Recommended Posts