ChayutPiyaosotsun
commented
10 months ago @pavaris-pm I'm glad to hear from you and it's great to know you checked out our SETalyze repository. Thanks for your interest and appreciation!
Regarding your first question about the use of models from the Darts library, we've selectively utilized three models based on their capability to support past covariate in multivariate time series. Specifically, we're working with BlockRNN, Transformer, and N-BEATS for grid search and fine-tuning processes. Our choice was driven by the specific needs of our dataset and the unique strengths of these models.
As for the benchmarking query, I appreciate your curiosity about the performance of Darts in comparison to other models. To be transparent, I recently realized a few oversights in the initial modeling approach. However, we employed three metrics for evaluation: directional accuracy, MAPE, and RMSE. Our directional accuracy averages between 35-45%, while MAPE and RMSE are around 5-7%.
It's important to note that we initially built a singular model applicable across all stocks, which might not have yielded the most accurate benchmarks. From that recognition, our current focus is on developing individual models for each stock to enhance accuracy and reliability. Additionally, we're exploring integrating a more diverse set of algorithms to refine our approach further.
Thank you again krub for your interest.
Hi guys, hope this issue finds you well krub. we've a quick discussion at
Captial Market Datathonyesterday if you can remember.However, i already take a look at this
SETalyzerepo, handling time series dataset using Darts library is very impressive. Did you guys use all models in the library itself e.g. transformers-based time series forecasting model or others ?Apart from that, I want to ask that did you guys have any benchmark (maybe an evaluation result) between using darts compared to other models ? since i want to know how great it is and what is the performance of them on SET dataset krub.