|
1 ano
experiência
|
0
produtos
|
0
versão demo
|
|
1
trabalhos
|
0
sinais
|
0
assinantes
|
Computer Science Bachelor in CUHK(SZ)
Quant Researcher with 3+ years of trading experience
Currently managing 5+ trading systems
Specializes in CTA strategy development
Github: https://github.com/CodyOutcast
Quant Researcher with 3+ years of trading experience
Currently managing 5+ trading systems
Specializes in CTA strategy development
Github: https://github.com/CodyOutcast
Zhuo Kai Chen
I personally have some critical thoughts about developing machine learning models as filters for trend-following strategies. We all know that trend-following strategies primarily profit from a few outlier trades that offset most of the losses. This characteristic of profit distribution is difficult to capture with a binary classifier. While we can attempt to minimize this issue by assigning greater weight to the higher profit class, it remains challenging. Intuitively, predicting long-term profits is akin to forecasting prices, which academia often regards as a mystery. Dr. Ernest P. Chan, the author of "Quantitative Trading", stated that using tree models to predict short-term prices is much easier than predicting long-term prices—similar to how forecasting the weather for the next minute is easier than predicting it for tomorrow. I strongly agree and have found success using such models to predict short-term mean reversion strategies.
Recently, a fund manager from Man Group gave a lecture about CTAs (Commodity Trading Advisors) at my university. He mentioned that they rarely use machine learning in their CTA bots, which baffled me. Literally, one of the most successful firms in the world prefers simple rules and intuitive algorithms over sophisticated methods. I asked him why, and he explained:
1. They tried using machine learning to mine alphas but failed miserably.
2. They attempted to use it as a filter, similar to what we discussed in this article, but it barely worked, achieving only 80% correlation. This means it provided almost no additional edge compared to the original strategy.
3. They found success in using machine learning to select the best strategy for a given market.
Regarding the third point, I wondered why they didn’t simply test each strategy for every market and compare the results. However, I assume they find it more efficient to cluster markets for certain strategies, especially since they trade over 6,000 assets. They believe the aforementioned theory explains their obstacles, as they primarily use trend-following strategies for their CTA bots.
Recently, a fund manager from Man Group gave a lecture about CTAs (Commodity Trading Advisors) at my university. He mentioned that they rarely use machine learning in their CTA bots, which baffled me. Literally, one of the most successful firms in the world prefers simple rules and intuitive algorithms over sophisticated methods. I asked him why, and he explained:
1. They tried using machine learning to mine alphas but failed miserably.
2. They attempted to use it as a filter, similar to what we discussed in this article, but it barely worked, achieving only 80% correlation. This means it provided almost no additional edge compared to the original strategy.
3. They found success in using machine learning to select the best strategy for a given market.
Regarding the third point, I wondered why they didn’t simply test each strategy for every market and compare the results. However, I assume they find it more efficient to cluster markets for certain strategies, especially since they trade over 6,000 assets. They believe the aforementioned theory explains their obstacles, as they primarily use trend-following strategies for their CTA bots.
· 4
Zhuo Kai Chen
Publicado o artigo Utilizando o modelo de Machine Learning CatBoost como Filtro para Estratégias de Seguimento de Tendência 
CatBoost é um poderoso modelo de machine learning baseado em árvores que se especializa em tomada de decisão com base em features estacionárias. Outros modelos baseados em árvores como XGBoost e Random Forest compartilham características semelhantes em termos de robustez, capacidade de lidar com padrões complexos e interpretabilidade. Esses modelos têm uma ampla gama de usos, desde análise de features até gestão de risco. Neste artigo, vamos percorrer o procedimento de utilização de um modelo CatBoost treinado como filtro para uma estratégia clássica de seguimento de tendência com cruzamento de médias móveis.
· 2
231