High-frequency factors continue to be updated, and the gating model once again strengthens the factor effects!
In the first half of 2024, Datayes constructed two factor libraries with rich dimensionality and strong signal characteristics using exchange high-frequency quotes that were authoritatively licensed by the exchanges. Both contain 153 single-day high-frequency quotes with the original physical characteristics of the feature factors carved, and also contain T-20 smoothed 121 high-frequency quotes factors.
The two sets of factor libraries on the new launch of multi-user testing down to harvest less than a good feedback. We continue to research and plan to add model training on the basis of a single factor to obtain a better factor.
At present, we launched the first synthetic factor: hist_x453_yspe20d
The factor construction goes like this:
1) A 30-dimensional feature factor obtained by using daily frequency high open low close volume turnover vwap, rolling the average over the past N days
2) Using 141 daily frequency eigenfactors, the 20-day mean, standard deviation, and current value are computed, with an initial 423-dimensional factor
3) Introduce the characteristics of the market index (the past 5/10/20/60 days return, volume ratio), gating model according to the market index, from the 423-dimensional selection of 54-dimensional features, and the first step of the fixed 30-dimensional splicing
4) Forecast target: the next 20 days trait return (removing 10 styles and industries)
Tested down the IC and grouping effect to get a pretty good situation, welcome to check, interested in contacting us！