Prophet is a Facebook Research project launched in February 2017. In another post, we have been playing with it by using financial time series. In this time we compare the produced forecast with the actual real price in our chosen interval.
“Prediction is very difficult, especially about the future.”
Last March we published a first approach to “Prophet” – the time series forecasting tool developed by Facebook Research. We can choose the training and forecast time windows, and produce the price forecast of any Yahoo Finance symbol.
Now we move forward another step by comparing the forecast produced by “Prophet” with the actual real price in our chosen interval (link to the app). You can choose the initial date for the training period (ending December 30th, 2016). We forecast the price for the period between January 1st to June 23rd 2017.
In some cases, the forecast matches the real price path very well, but in other cases they move in opposite directions (the prices for the forecast period are coloured accordingly).
Have you used the “Prophet” tool yet? Do you have any comments about our tests? Write to us below!
With credit to:
Sean J. Taylor, Ben Letham and the Facebook Research Team.
Code:
ui.r
fluidPage(
titlePanel("Playing with Prophet on Financial Time Series (II)"),
fluidRow(
column(3, wellPanel(
h3("\"Prediction is very difficult, especially about the future.\""),
helpText("Prophet is a procedure for forecasting time series data developed by Facebook's Core Data Science team."),
helpText("Using this app you can compare de Prophet prediction for financial time series (from Yahoo Finance) with the real prices"),
helpText("You can choose the initial date for the train period (ending December 30th, 2016). We forecast price for June 23rd, 2017 (Green color same direction, red color different)."),
dateInput("date", "Initial Date:", "2014-01-01", max="2016-01-01"),
textInput("sym", "Symbol (Yahoo Finance!)", "AAPL"),
helpText("Nice Examples: TLT (4 years), AMZN (5), BND (3)"),
helpText("Ugly Examples: XOM (3 years), CVX (4), NKE (5)"),
checkboxInput("seasonal","Use yearly seasonal factor:", TRUE),
h3("Links"),
p(tags$a(href="https://mydata.shinyapps.io/ShinyProphetV2/", "Playing with Prophet on Financial Time Series")),
p(tags$a(href="https://research.fb.com/prophet-forecasting-at-scale", "Research Facebook Prophet")),
p(tags$a(href="https://facebookincubator.github.io/prophet", "Github")),
p(tags$a(href="https://cran.r-project.org/web/packages/prophet/index.html", "Prophet R Package")),
helpText("Data from Yahoo Finance!")
)),
column(6,
plotOutput("plot1", width = "1217px", height = "800px")
)
)
)
server.r
library(prophet)
library(quantmod)
library(data.table)
library(ggplot2)
library(dplyr)
function(input, output) {
output$plot1 <- renderPlot({
mydf <- getSymbols(input$sym, src = "yahoo", from = input$date, to = "2016-12-30", auto.assign = FALSE)
mydf <- data.frame(mydf[,6])
setDT(mydf, keep.rownames = TRUE)
colnames(mydf) <- c("ds", "y")
m <- prophet(mydf, yearly.seasonality = input$seasonal, weekly.seasonality = FALSE)
future <- make_future_dataframe(m, periods = 177)
forecast <- predict(m, future)
mydf2 <- getSymbols(input$sym, src = "yahoo", from = input$date, to = "2017-06-23", auto.assign = FALSE)
mydf2 <- data.frame(mydf2[,6])
colnames(mydf2) <- "Price"
mydf2$ds <- rownames(mydf2)
rownames(mydf2) <- NULL
mydf2$ds <- as.Date(mydf2$ds)
perry = left_join(mydf2, forecast, by="ds")
rownames(perry) <- perry$ds
perry$PriceX <- perry$Price
perry$Price[perry$ds > "2016-12-30"] = NA
perry$PriceX[perry$ds < "2017-01-01"] = NA
priceref = perry$Price[perry$ds == "2016-12-30"]
yhatend = tail(perry$yhat,1)
priceend = tail(perry$PriceX,1)
theme_set(theme_gray(base_size = 18))
if (((yhatend>priceref) && (priceend>priceref)) | ((yhatend<priceref) && (priceend<priceref))) {
ggplot(perry, aes(x=ds)) + geom_point(aes(y=Price)) + geom_point(aes(y=PriceX), col="forestgreen", size=3) +
geom_line(aes(y=yhat)) + geom_ribbon(aes(ymin=yhat_lower, ymax=yhat_upper), fill="deepskyblue", alpha=0.5) +
geom_line(aes(y=priceref), cex=0.8, col="steelblue4") +
geom_vline(aes(xintercept=as.numeric(as.Date("2017-01-01"))), cex=0.8, col="steelblue4")}
else {
ggplot(perry, aes(x=ds)) + geom_point(aes(y=Price)) + geom_point(aes(y=PriceX), col="firebrick1", size=3) +
geom_line(aes(y=yhat)) + geom_ribbon(aes(ymin=yhat_lower, ymax=yhat_upper), fill="deepskyblue", alpha=0.5) +
geom_line(aes(y=priceref), cex=0.8, col="steelblue4") +
geom_vline(aes(xintercept=as.numeric(as.Date("2017-01-01"))), cex=0.8, col="steelblue4")}
})
}
related posts
0 Comments