ahaeusser
commented
3 years ago Hello swaheera,
thank you very much for your feedback and your interest into the echos package.
I took a closer look at your code and made a few adjustments. In some places I wasn't sure what exactly you want to do, so I just ran through the code the way I think it should be done.
Notes to your case:
- The case below is created for monthy time series data. I was not sure about the desired frequency of your time series, so I assumed monthly frequency
- The time series cross-validation should be adjusted carefully to your time series data and I made some assumptions (the values from the README are for hourly time series data)
- The example below trains and forecasts the model only for one split - this is just for testing purposes.
- The error message appears because the arguments in ESN() were changed from upper case to lower case ("bic" instead of "BIC"). I haven't had time to adjust the README yet - sorry for that!
And here's the adjusted code:
# Load packages
library(echos)
library(tscv)
library(dplyr)
library(tsibble)
library(fabletools)
library(fable)
Sys.setlocale("LC_TIME", "C")
# Create artificial daily time series data
date <- seq(
from = as.Date("1800-01-01"),
to = as.Date("2020-01-01"),
by = "day")
damage <- rnorm(
n = 80354,
mean = 100,
sd = 10)
# Prepare daily time series data as tsibble
data <- tsibble(
date = date,
damage = damage,
index = date) %>%
pivot_longer(
cols = damage,
names_to = "Variable",
values_to = "Value")
# Aggregate daily to monthly time series data
data <- data %>%
group_by_key() %>%
index_by(year_month = ~ yearmonth(.)) %>%
summarise(Value = sum(Value, na.rm = TRUE)) %>%
rename(Date = year_month) %>%
select(Date, Variable, Value)
data
# Example proceeds with monthly time series data
# Setup for time series cross validation
n_init <- 200 # size for training window
n_ahead <- 12 # size for testing window (i.e. 12 month ahead forecast)
mode <- "slide" # fixed window approach (for such a long time series this is fine)
n_skip <- 11 # skip 11 month between windows
n_lag <- 0 # no lag
data <- data %>%
tscv::split_data(
n_init = n_init,
n_ahead = n_ahead,
mode = mode,
n_skip = n_skip,
n_lag = n_lag)
# For this example, only the first split is considered
data <- data %>%
filter(split == 1)
models <- data %>%
filter(sample == "train") %>% # only training data should be used for model training
model(
"ESN" = ESN(
Value,
inf_crit = "bic", # here's the error: the argument changed from upper case to lower case ("BIC" -> "bic") and forgot to adjust the README
max_lag = 3,
n_initial = 50,
n_res = 200,
scale_inputs = c(-1, 1)),
"Naive" = NAIVE(Value))
# Detailed report of ESN
models %>%
select(ESN) %>%
report()
# Forecast models
fcsts <- models %>%
forecast(h = n_ahead)
fcsts
# Visualize the forecast and the actual values
actuals <- data %>%
filter_index("1816 Jan" ~ .)
fcsts %>%
autoplot(
actuals,
level = NULL,
size = 1)I hope I could help you a little bit with the code above.
Disclaimer: The echos package is highly experimental and it is very likely that there will be (substantial) changes in the near future. These changes will probably affect the interface (e.g. arguments within ESN()) and the underlying modeling procedure. Just as a friendly warning ;)
I wish you a Merry Christmas and a Happy New Year!
Best, Alex
swaheera
Hello, great job creating this package for echo state networks in R!
I an trying to use your code for a univariate time series I generated and I am running into some problems relating to "missing values not supported". Could you please take a look at this if you have time?
Here is the error:
I tried doing as the error message suggested:
But the same error happens. This error is preventing me from running the rest of the code:
Could you please take a look at this if you have time? Perhaps I am doing something wrong and you could kindly show me how to fix this problem?
Thanks you so much!