How to improve results accuracy for time series forecast?

Question

I would like to share my analysis in order to have guidance on how to improve the time series results.

Here you will find a table comparing real values vs forecasted ones.

Below you will find the code that generated that output (in R).

# Load required libraries
library(forecast)
library(lubridate)
library(tidyverse)
library(scales)
library(ggfortify)

# Load dataset
emea <- read.csv(file="C:/Users/nsoria/Google Drive/Trabajo/AMS Globales/812_Finanzas.csv", header=TRUE, sep=';', dec=",")

# Create time series object
ts_fin <- ts(emea$Valor, deltat = 1/24, start = c(2015, 1))

# Pull out the seasonal, trend, and irregular components from the time series 
model <- stl(ts_fin, s.window = "periodic")

# Predict the next 3 bi weeks of tickets
pred <- forecast(model, h = 5)

# Round values to better accuracy
pred$mean <- round(pred$mean)
pred$upper <- round(pred$upper)
pred$lower <- round(pred$lower)

I have added in this link the dataset which is used here.

My main concern would be how to improve accuracy for forecasted results. I am looking into a possibility to add a machine learning algorithm but I am open to suggestions.

Thanks in advance.

Edit 24/01: As suggested, you will find the actual results for the prediction and the different outputs.

Answer 1

Without looking at your data, it seems that the only method you have used is stl().

There are other models available in the forecast package, including ets() (exponential smoothing methods), auto.arima() (arima models), nnetar() (neural networks).

That would be the first thing to try to see if your can improve your forecasts.

Answer 2

Also, try plotting the series. Check how the trend looks like using decompose methods. Check for the stationarity of the series with Dickey-Fuller test or KPSS test. Use auto.arima to find p,d,q or try finding with acf, pacf plots. You have missed a lot.