• arima
• construction
• data analysis
• ensamble
• forecast
• forecaster
• metrics
• model-agnostic
• models
• neural networks
• predictions
• probability
• statistics
• time series
• trees
• weights

Mathematically speaking, time series are sets of observations that are generated sequentially over time.
In this work, many different models that can describe time series have been studied, such as ARIMA models, exponential smoothing models, LSTM neural networks and tree-based models: Random Forests, XGBoost and LightGBM.
The aim of this thesis is to create a forecasting model for time series that can be used by an inexpert user. The model has to be, firstly, completely automated and, secondly, its behaviour has to properly adapt to each time series that will be processed, that is in a model-agnostic way.
The idea behind this process is to test the time series models cited above on the data and then to combine these
models, obtaining a new meta-model. In order to produce the best forecasts, two ways in which these predictions can be combined have been investigated. The first one consists of choosing the best individual predictions as the final forecast. The second one consists of weighing the predictions in order to emphasize the best guesses. For doing so, a stepwise greedy method has been implemented.
Moreover, how to automatically evaluate the predictions has been analyzed. To do so, a specific rank-based evaluation system has been developed. The final forecast is identified as the one with the top score for this ranking system.
Several experiments have been carried out, in order to test the accuracy and the case-specific results of the model-agnostic forecaster.