gpt4 book ai didi

R拟合和预测每日时间序列

转载 作者:行者123 更新时间:2023-12-05 07:46:57 27 4
gpt4 key购买 nike

我正在处理每日时间序列,我需要根据我的历史记录构建 90 天(或更多)的预测 - 当前时间序列大约有 298 个数据点。

我遇到的问题是最终预测中著名的平线 - 是的,我可能没有季节性,但我正在努力解决这个问题。另一个问题是如何找到最佳模型并从现在开始针对这种行为进行调整。

我创建了一个测试用例来进一步调查此问题,我们将不胜感激。

谢谢,

开始

x <- day_data  # My time serie
z <- 90 # Days to forecast

low_bound_date <- as.POSIXlt(min(x$time), format = "%m/%d/%Y") # oldest date in the DF.

> low_bound_date
[1] "2015-12-21 PST"

low_bound_date$yday
> low_bound_date$yday # Day in Julian
[1] 354

lbyear <- as.numeric(substr(low_bound_date, 1, 4))
> lbyear
[1] 2015

这是我的时间系列内容

> ts
Time Series:
Start = c(2065, 4)
End = c(2107, 7)
Frequency = 7
[2] 20.73 26.19 27.51 26.11 26.28 27.58 26.84 27.00 26.30 28.75 28.43 39.03 41.36 45.42 44.80 45.33 47.79 44.70 45.17
[20] 34.90 32.54 32.75 33.35 34.76 34.11 33.59 33.60 38.08 30.45 29.66 31.09 31.36 31.96 29.30 30.04 30.85 31.13 25.09
[39] 17.88 23.73 25.31 31.30 35.18 34.13 34.96 35.12 27.36 38.33 38.59 38.14 38.54 41.72 37.15 35.92 37.37 32.39 30.64
[58] 30.57 30.66 31.16 31.50 30.68 32.21 32.27 32.55 33.61 34.80 33.53 33.09 20.90 6.91 7.82 15.78 7.25 6.19 6.38
[77] 38.06 39.82 35.53 38.63 41.91 39.76 37.26 38.79 37.74 35.61 39.70 35.79 35.36 29.63 22.07 35.39 35.99 37.35 38.82
[96] 25.80 21.31 18.85 9.52 20.75 36.83 44.12 37.79 34.45 36.05 16.39 21.84 31.39 34.26 31.50 30.87 28.88 42.83 41.52
[115] 42.34 47.35 44.47 44.10 44.49 26.89 18.17 40.44 43.93 41.56 39.98 40.31 40.59 40.17 40.22 40.50 32.68 35.89 36.06
[134] 34.30 22.67 12.56 13.29 12.34 28.00 35.27 36.57 33.78 32.15 33.58 34.62 30.96 32.06 33.05 30.66 32.47 30.42 32.83
[153] 31.74 29.39 22.39 12.58 16.46 5.36 4.01 15.32 32.79 31.66 32.02 27.60 31.47 31.61 34.96 27.77 31.91 33.94 33.43
[172] 26.94 28.38 21.42 24.51 23.82 31.71 26.64 27.96 29.29 29.25 28.70 27.02 27.62 30.90 27.46 27.37 26.46 27.77 13.61
[191] 5.87 12.18 5.68 4.15 4.35 4.42 16.42 25.18 26.06 27.39 27.57 28.86 15.18 5.19 5.61 8.28 7.78 5.13 4.90
[210] 5.02 5.27 16.31 25.01 26.19 25.96 24.93 25.53 25.56 26.39 26.80 26.73 26.00 25.61 25.90 25.89 13.80 6.66 6.41
[229] 5.28 5.64 5.71 5.38 5.76 7.20 7.27 5.55 5.31 5.94 5.75 5.93 5.77 6.57 5.52 5.51 5.47 5.69 19.75
[248] 29.22 30.75 29.63 30.49 29.48 31.83 30.42 29.27 30.40 29.91 32.00 30.09 28.93 14.54 7.75 5.63 17.17 22.27 24.93
[267] 35.94 37.42 33.13 25.88 24.27 37.64 37.42 38.33 35.20 21.32 7.32 4.81 5.17 17.49 23.77 23.36 27.60 26.53 24.99
[286] 24.22 23.76 24.10 24.22 27.06 25.53 23.40 37.07 26.52 25.19 28.02 28.53 26.67

第一步,我在 ts 中获取我的数据

day_data_ts <- ts(x$avg_day, start = c(lbyear,low_bound_date$yday), frequency=7)

plot(day_data_ts)

plot_ts

acf(day_data_ts)

acf_ts

第二步,我在 msts 中获取我的数据

day_data_msts <- msts(x$avg_day, seasonal.periods=c(7,365.25), start = c(lbyear,low_bound_date$yday))

plot(day_data_msts)

acf(day_data_msts)

我进行了几次拟合迭代,试图找出最佳拟合和预测模型。

第一个拟合测试仅使用 ts

fit1 <- HoltWinters(day_data_ts)
> fit1
Holt-Winters exponential smoothing with trend and additive seasonal component.
Call: HoltWinters(x = day_data_ts)
Smoothing parameters: alpha: 1 beta : 0.006757112 gamma: 0

Coefficients:
[,1]
a 28.0922449
b 0.1652477
s1 0.6241837
s2 1.9084694
s3 0.9913265
s4 0.8198980
s5 -1.7015306
s6 -1.2201020
s7 -1.4222449


fit2 <- tbats(day_data_ts)
> fit2
BATS(1, {0,0}, 0.8, -)
Parameters: Alpha: 1.309966 Beta: -0.3011143 Damping Parameter: 0.800001
Seed States:
[,1]
[1,] 15.282259
[2,] 2.177787
Sigma: 5.501356 AIC: 2723.911


fit3 <- ets(day_data_ts)
> fit3
ETS(A,N,N)
Smoothing parameters: alpha = 0.9999
Initial states: l = 25.2275
sigma: 5.8506
AIC AICc BIC
2756.597 2756.678 2767.688


fit4 <- auto.arima(day_data_ts)
> fit4
ARIMA(1,1,2)
Coefficients:
ar1 ma1 ma2
0.7396 -0.6897 -0.2769
s.e. 0.0545 0.0690 0.0621
sigma^2 estimated as 30.47: log likelihood=-927.9
AIC=1863.81 AICc=1863.94 BIC=1878.58

第二个测试使用 msts。我还将 ets 模型更改为 MAM

fit5 <- tbats(day_data_msts)
> fit5
BATS(1, {0,0}, 0.8, -)
Parameters: Alpha: 1.309966 Beta: -0.3011143 Damping Parameter: 0.800001
Seed States:
[,1]
[1,] 15.282259
[2,] 2.177787
Sigma: 5.501356 AIC: 2723.911


fit6 <- ets(day_data_msts, model="MAN")
> fit6
ETS(M,A,N)
Smoothing parameters: alpha = 0.9999 beta = 9e-04
Initial states: l = 52.8658 b = 3.9184
sigma: 0.3459
AIC AICc BIC
3042.744 3042.949 3061.229


fit7 <- auto.arima(day_data_msts)
> fit7
ARIMA(1,1,2)
Coefficients:
ar1 ma1 ma2
0.7396 -0.6897 -0.2769
s.e. 0.0545 0.0690 0.0621
sigma^2 estimated as 30.47: log likelihood=-927.9
AIC=1863.81 AICc=1863.94 BIC=1878.58

最佳答案

您可以按如下方式对先前估计的模型进行预测(使用内置时间序列 LakeHuron):

library(forecast)
y <- LakeHuron
tsdisplay(y)
# estimate ARMA(1,1)
mod_2 <- Arima(y, order = c(1, 0, 1))
#make forecast for 5 periods (years in this case)
fHuron <- forecast(mod_2, h = 5)
#show results in table
fHuron
#plot results
plot(fHuron)

这会给你: enter image description here请注意,ARIMA 模型的预测基于先前的值,因此如果我们对多个周期进行预测,该模型将使用已经预测的值来预测下一个。这会降低准确性。

要拟合最佳 ARIMA 模型,请使用此函数:

library(R.utils) #for the function 'withTimeout'
fitARIMA<-function(timeseriesObject, timout)
{
final.aic <- Inf
final.order <- c(0,0,0)
for (p in 0:5) for (q in 0:5) {
if ( p == 0 && q == 0) {
next
}

arimaFit = tryCatch(
withTimeout(arima(timeseriesObject
,order=c(p, 0, q))
,timeout = timeout)
,error=function( err ) FALSE
,warning=function( err ) FALSE )

if( !is.logical( arimaFit ) ) {
current.aic <- AIC(arimaFit)
if (current.aic < final.aic) {
final.aic <- current.aic
final.order <- c(p, 0, q)
final.arima <- arima(timeseriesObject, order=final.order)
}
} else {
next
}
}
final.order<-c(final.order,final.aic)
final.order
}

关于R拟合和预测每日时间序列,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/40387365/

27 4 0
Copyright 2021 - 2024 cfsdn All Rights Reserved 蜀ICP备2022000587号
广告合作:1813099741@qq.com 6ren.com