[PYTHON] Dereste event attack format 1st place point estimation -State space model using Pystan-

Overview

I learned how to analyze a state-space model using Stan in a book, so I practiced it with actual data.
The goal is "estimate the first point at the end of the event at the start date of the event"
This time, I used the "local linear trend" + "time variable" model.
** As a result, it seems that a model with sufficient accuracy was created **
The first place is "trend" and "length of event period", which I could almost explain.
After all, it was easier than Prediction of 2001
It seems that the "excitement of each event" does not affect the 1st place so much, and the apts run for the full time with the highest efficiency.
[Place here](https://github.com/kzy611/qiita/tree/master/dereste_analysis/dereste_event_point_ranking_border/%E3%82%A2%E3%82%BF%E3%83] % 9D% E3% 83% B3_1 / script).
Forecast of 2001 is difficult, and I am considering explanatory variables, so I tried it as a change of mood.

Data acquired by here extracted only from the format of Attapon

[See Kochi for formulas](https://qiita.com/kazuya_minakuchi/items/09b010927688b322df9d#%E3%83%AD%E3%83%BC%E3%] 82% AB% E3% 83% AB% E7% B7% 9A% E5% BD% A2% E3% 83% 88% E3% 83% AC% E3% 83% B3% E3% 83% 89% E3% 83% A2% E3% 83% 87% E3% 83% AB)
Red line is point estimation, red range is interval estimation (5% -95% interval)
The correct answer rate for interval estimation for the data used to create the model is 91.7% (total number: 60, number of correct answers: 55).

Comparing the trend component and the coefficient, there is a tendency that "trend is smooth" and "coefficient changes suddenly".
Factors that affect time efficiency, such as "addition of high-performance characters" and "implementation of Grand Drive", seem to have been added to the coefficient.

Let's look at the ratio as "ratio = measured value / predicted value"
The idea that "the larger the value, the larger the deviation tends to be."
Perfect prize for "ratio = 1"
Predicted value <measured value when 1 or more. Measured value <predicted value below -1

The average is about 1 (this is natural because the model is made aiming at that)
With standard deviation 0.043 = normal distribution, 68% falls within ± 4.3%
Maximum deviation is about 13% (min, max)
50% is less than 3.3% (25% -75%)

I also looked at them in descending order and by month, but there was not much regularity.

To create the model of ↑, [Data before Hero Versus Reinanjo in the event list](https://imascg-slstage-wiki.gamerch.com/%E3%82%A4%E3%83%99%E3 % 83% B3% E3% 83% 88% E3% 83% 87% E3% 83% BC% E3% 82% BF), so the next [Orange Time](https: // imascg-) slstage-wiki.gamerch.com/%E3%80%90%E3%82%A4%E3%83%99%E3%83%B3%E3%83%88%E3%80%91%E3%82%AA I tried to estimate the 1st place point of% E3% 83% AC% E3% 83% B3% E3% 82% B8% E3% 82% BF% E3% 82% A4% E3% 83% A0)
"Orange Time" information
Date: 2020/09/20
Period: 174h
1st place point (answer): ** 1,250,000 **
Predicted value
Point estimation: ** 1,222,458 **
Interval estimation (90%): ** 1,102,110 ~ 1,344,647 **
The actual value is included in the interval estimation, and the estimation is successful.