Thesis Stuff/Notes/TODOs

This is a repo dedicated to my MSc Thesis

Abstract (to be filled ...)

Instructions

• Clone the repository using

git clone https://github.com/alexliap/fts_explore.git

• to be filled ...

✅ TODO

• Find univariate dataset for first tests

• Finetune MOIRAI on this dataset in order to beat on train and validation

• Find a univariate dataset of a subdomain of the first one and test thesis hypothesis

  • Make iterative train loop
  • Test 3 different configurations for finetuning:
    • Use model of training iteration N-1 for training iteration N. 1 backward pass.
    • At each training iteration use the Stage 1 model for finetuning. 1 backward pass.
    • At each training iteration use the Stage 1 model for finetuning. Multiple backward passes.
      • Without dropout
      • With 10% dropout
      • With 20% dropout
  • For Stage 2, also perform iterative training of the pretrained model in order to test the hypothesis
    • Without dropout
    • With 20% dropout
  • Experiment with different ways of evaluation/visualization
    • Get forecasts and targets, in order to calculate whichever metric

• Create pipeline where all the above steps are excecuted with the use of Hydra configuration files.

  • Create pipeline for Stage 1
  • Create pipeline for Stage 2
  • Test pipeline end to end

• Search for other domains & subdomains to make experiments

  • Daily crpto data (BTC & ETH)
    • Split train/validation data to 2022-01-01 => 1004 data points for validation for Stage 1
  • Weather hourly & daily data from Open Meteo (Athens & Smunri)
    • Daily: Train 2014-2021 | Validation 2021-2024

• Add WQL as a comparison metric. WQL stands for weight quantile loss and is implemented by gluonTS as MeanWeightedSumQuantileLoss. It says mean beacuse it computes WeightedSumQuantileLoss for several quantiles and then calcualtes the mean.