Stats › Figures exploitables › Text models
A word list captures a third of what an estimated text model does, and costs eat most of both
| Method and weighting | Ratio de Sharpe | Basis points a day |
|---|---|---|
| Estimated model, equal weighted | 4.29 | 33 |
| Commercial feed, equal weighted | 3.24 | 18 |
| Standard finance word list, equal weighted | 1.71 | 12 |
| Estimated model, value weighted | 1.33 | 10 |
| Commercial feed, value weighted | 1.14 | 8 |
| Standard finance word list, value weighted | 0.73 | 5 |
Out of sample daily long short performance on 31.5 million newswire articles from 1 January 1989 to 31 July 2017, of which 10.4 million are tagged to a single firm. Rolling fifteen year training windows with one year out of sample, repeated fourteen times. Note the value weighted column: every method loses between two thirds and three quarters of its Sharpe ratio when the portfolio is weighted by company size.
On the same newswire corpus, an estimated sentiment model gives an out of sample Sharpe ratio of 4.29 and a standard finance word list gives 1.71. At ten basis points a day of cost the best achievable net Sharpe falls to 2.30.
Réservé aux membres
La suite de cette page est réservée aux membres
Below this point there are 5 sections and 3 named sources, roughly 750 words of it. Every figure carries the source it came from and the date the data is from.
Vous pouvez continuer à parcourir gratuitement toutes les statistiques de la bibliothèque. L'introduction et le résumé sont toujours accessibles.