I am doing forecasting with auto.Arima with uni-variate data but my forecast is not correct. I have used all the steps correctly but the point forecast value is not coming out to be right. Please help me.
Here is my data:
s <- read.csv(url('https://ondemand.websol.barchart.com/getHistory.csv?apikey=c3122f072488a29c5279680b9a2cf88e&symbol=zs*1&type=dailyNearest&backAdjust=false&startDate=20100201'))
Here is my code:
data <- s[c(3, 7)]
summary(data)
data1.ts <- zoo(data[,2], seq(from = as.Date("2010-02-01"), to = as.Date("2022-05-13"), by = 1))
autoplot(data1.ts)
Arima Model:
fit_arima <- auto.arima(data1.ts, stepwise = FALSE, approximation = FALSE, trace = TRUE)
print(summary(fit_arima))
checkresiduals(fit_arima)
forecast_Arima <- forecast(fit_arima, h = 1)
forecast_Arima
Foreacst Value:
Point Forecast Lo 80 Hi 80 Lo 95 Hi 95
19126 976.4357 949.874 1002.997 935.813 1017.058
Little update:
I have tried to load the data as a ts object and have got the accurate Point forecast value but, my forecast year is not correct. The one-step-ahead forecasting is giving me value for the year 2021 but my end date is 2022-05-13. I just want to correct the year. This is new code:
ts_soy <- ts(data[,2], start = c(2010-02-01), frequency = 214)
autoplot(ts_soy)
fit_arima <- auto.arima(ts_soy)
print(summary(fit_arima))
checkresiduals(fit_arima)
forecast_Arima <- forecast(fit_arima, h = 1)
forecast_Arima
Point Forecast Lo 80 Hi 80 Lo 95 Hi 95
2021.472 1646.5 1625.071 1667.929 1613.727 1679.273
CodePudding user response:
I can reproduce your issue, and the reason is that your data1.ts
contains too much data. You are trying to get rid of the weekends to create a continuous timeseries (aka timeseries without gaps). The principle is correct but you are exceeding the amount of records you have by 1388 records. Since R tends to recycle values, you get closing prices from the early years again and this influences the arima
function.
You can do something like create a timeseries starting from the earliest date and to is this date the number of records - 1
data.ts <- zoo(data[,2], seq(from = as.Date("2010-02-01"),
to = as.Date("2010-02-01") 3096,
by = 1))
fit_ar <- forecast::auto.arima(data.ts, stepwise = FALSE, approximation = FALSE)
forecast::forecast(fit_ar, h = 1)
Point Forecast Lo 80 Hi 80 Lo 95 Hi 95
17738 1648.129 1626.759 1669.499 1615.446 1680.812
This is one of the reasons why I prefer to use fable, I can inspect the data better.
library(fpp3)
library(fable.prophet)
fit <- data %>%
mutate(id = row_number()) %>% # create index to use otherwise timeseries has gaps
tsibble(index = id) %>%
model(naive = NAIVE(close),
arima = ARIMA(close, stepwise = FALSE, approximation = FALSE),
)
forecast(fit, h = 1)
# A fable: 2 x 4 [1]
# Key: .model [2]
.model id close .mean
<chr> <dbl> <dist> <dbl>
1 naive 3098 N(1646, 280) 1646.
2 arima 3098 N(1649, 278) 1649.
# prophet needs dates and can handle weekends
prophet_fit <- data %>%
mutate(tradingDay = ymd(tradingDay)) %>%
tsibble() %>%
model(prophet_model = prophet(close))
forecast(prophet_fit, h = 1)
# A fable: 1 x 4 [1D]
# Key: .model [1]
.model tradingDay close .mean
<chr> <date> <dist> <dbl>
1 prophet_model 2022-05-14 sample[5000] 1657.