Uloženo v:
| Hlavní autor: | |
|---|---|
| Médium: | Recurso digital |
| Jazyk: | |
| Vydáno: |
Zenodo
2025
|
| Témata: | |
| On-line přístup: | https://doi.org/10.5281/zenodo.15327832 |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
Obsah:
- <p>This paper proposes an innovative Transformer model, Single-directional representative<br>from Transformer (SERT), for US large capital stock pricing. It<br>also innovatively applies the pre-trained Transformer models under the stock<br>pricing and factor investment context. They are compared with standard Transformer<br>models and encoder-only Transformer models in three periods covering<br>the entire COVID-19 pandemic to examine the model adaptivity and suitability<br>during the extreme market fluctuations. Namely, pre-COVID-19 period (mild<br>up-trend), COVID-19 period (sharp up-trend with deep down shock) and 1-year<br>post-COVID-19 (high fluctuation sideways movement). The best proposed SERT<br>model achieves the highest out-of-sample R2, 11.2% and 10.91% respectively,<br>when extreme market fluctuation takes place followed by pre-trained Transformer<br>models (10.38% and 9.15%). Their Trend-following-based strategy wise performance<br>also proves their excellent capability for hedging downside risks during<br>market shocks. The proposed SERT model achieves a Sortino ratio 47% higher<br>than the buy-and-hold benchmark in the equal-weighted portfolio and 28% higher<br>in the value-weighted portfolio when the pandemic period is attended. It proves<br>that Transformer models have a great capability to capture patterns of temporal<br>sparsity data in the asset pricing factor model, especially with considerable<br>volatilities. We also find the softmax signal filter as the common configuration<br>of Transformer models in alternative contexts, which only eliminates differences<br>between models, but does not improve strategy-wise performance, while increasing<br>attention heads improve the model performance insignificantly and applying<br>the ’layer norm first’ method do not boost the model performance in our case.</p>