我2个月大预测概念,但我正在努力自己学习和不断练习。这里我试图预测每周的产品在训练数据集上使用不同的运动预测技术,并在测试数据集上测试其准确性。我尝试过不同的技术,如ARIMA, TBATS, Holts Winter, UCM,贝叶斯结构时间序列等。但不能提高我的准确性。准确性似乎很糟糕。不知道我哪里做错了。我也尝试了ARIMA与回归,但它再次没有帮助我太多。我不确定是我的代码还是我的方法错了。谁能指导我提高我的准确性?以下是每周数据集(从2012年12月8日开始)
[1] 74 76 78 63 58 58 57 56 85 73 71 91 85
[14] 79 101 74 86 98 131 90 127 116 320 145 121 148
[27] 112 141 153 118 151 151 152 90 147 123 266 99 110
[40] 146 134 76 81 100 80 323 15 22 14 13 19 56
[53] 78 79 70 79 24 26 31 35 45 33 41 41 61
[66] 91 83 76 57 68 87 82 105 76 107 116 105 124
[79] 127 149 124 120 111 122 134 87 80 81 89 40 63
[92] 112 85 131 97 51 65 74 70 47 62 60 49 47
[105] 56 64 57 58 45 56 60 49 82 49 61 71 61
[118] 92 90 75 69 114 79 144 121 133 132 114 124 152
[131] 125 112 128 124 152 95 64 59 91 132 146 120 196
[144] 212 115 125 66 68 78 83 74 300 46 98 86 95
[157] 61 73 89 56 81 60 58 101 482 55 124 72 57
[170] 51 82 55 68 105 153 113 105 85 34 77 95 96
[183] 97 94 81 104 76 97 65 42 18 11
我认为我的培训期为178周,测试期为14周。假设'data'是我的dataframe以' units '作为我的colname
series <- ts(data, start=2012+342/365.25, frequency = 365.25/7)
kk <- 178
seas <- 365.25/7
st <- tsp(series)[1] + (1/seas)*(kk-1)
training <- window(series, end = st)
testing <- window(series, start = st + 1/52.17857, end = st+14/52.17857)
train1 <- training[,"units"]
test1 <- testing[,"units"]
##ARIMA
farima <- forecast(auto.arima(train1),h=14)
acc_arima <- accuracy(farima$mean,test1)
##TBATS
fTBATS <- forecast(tbats(train1,seasonal.periods=c(4,7,12,52)), h=14)
acc_TBATS <- accuracy(fTBATS$mean,test1)
##struTs
fstruTs <- forecast(StructTS(train1), h=14)
acc_struTs <- accuracy(fstruTs$mean,test1)
##UCM
forUCM <- ucm(formula = train1~0, data = train1, level =
TRUE, slope = TRUE)
fUCM <- predict(forUCM$model, n.ahead = 14)
acc_struTs <- accuracy(fUCM$fit,test1)
##Bayesian Structural time series
ss <- AddLocalLinearTrend(list(), train1)
ss <- AddSeasonal(ss, train1, nseasons = 52, season.duration = 7)
model2 <- bsts(train1, state.specification = ss, niter = 500)
fbsts <- predict(model2, horizon = 14, burn = 100)
acc_bsts <- accuracy(fbsts$mean,test1)
对于以上所有方法,我的MAPE都在100%以上,我认为这是非常糟糕的。有人能指导我提高准确性吗?我将不胜感激。谢谢!
我有几点建议:
1)如果你正在使用优秀的R forecast
包,我建议你至少尝试一下全自动预测(见下面的例子)。
2)我建议绘制预测值和实际值,以及历史数据,看看给定历史数据的输出是否合理。
3)我建议阅读一些R forecast
包的创建者制作的免费在线教科书。
下面的示例使用来自forecast
包的全自动时间序列预测,并绘制结果,包括您上面使用的数据集和另一个公开可用的数据集。
library(ggplot2)
library(forecast)
data <- read.table("./data.txt", quote=""", comment.char="")
series <- ts(as.numeric(data), start=2012+342/365.25, frequency = 365.25/7)
train_length <- 178
test_length <- length(series) - train_length
train_end <- time(series)[train_length]
test_start <- time(series)[train_length+1]
training <- window(series, end = train_end)
testing <- window(series, start = test_start)
## Use default forecast
fcast <- forecast(training, h=test_length)
plot(fcast)
lines(testing, col='red')
acc_fcast <- accuracy(fcast$mean, testing)
births <- scan("http://robjhyndman.com/tsdldata/data/nybirths.dat")
birthstimeseries <- ts(births, frequency=12, start=c(1946,1))
train_length <- 150
test_length <- length(birthstimeseries) - train_length
train_end <- time(birthstimeseries)[train_length]
test_start <- time(birthstimeseries)[train_length+1]
training <- window(birthstimeseries, end = train_end)
testing <- window(birthstimeseries, start = test_start)
## Use default forecast
fcast <- forecast(training, h=test_length)
plot(fcast)
lines(testing, col='red')
acc_births <- accuracy(fcast$mean, testing)