source: https://mp.weixin.qq.com/s/h4e8AOXCBxs_y8N-SKNkiA
Why ML could be used in finance
- Asset's risk premium is fundamentally a prediction problem; ML methods are largely specialized for prediction.
- Candidates of factors are large and highly correlated; ML is designed for dimensionality reduction and variable selection.
- Form in asset pricing is usually complex and unknown; ML methods are designed to approx complex nonlinear associations.
FRAMselect method
- finance factors are always highly correlated, 无法使用机器学习中常用的变量选择方法, 因为正则性不满足.
- 把变量提取成不同的因子: 个性和共性. 这些变量的相关性会很弱.
instruments for factor learning
- 如何利用instruments: 因子用instruments回归, 对拟合值做主成分分析.
裁剪与鲁棒性
- 我们用因子模型来解决观测到股票相关的问题。我们是用裁剪数据来解决鲁棒问题,对回归问题我们用裁剪损失函数,即Huber损失函数,来得到鲁棒性质的。我们说过预测的好特征必须是鲁棒。
HFT
- Momentum
- Momentums and durations are predicable.
文本数据与资产定价
- 学习文件和新闻的褒贬度,用它预测选择股票。传统一般用基于Dictionary的方法。
END
资产定价是一个预测问题, 机器可以学习金融。
1. 帮助处理大数据并从中选择重要因素和特征
2. 能很好应对过度拟合问题
3. 允许非线性学习极大地改善预测效果
4. 将稳健性和对抗性网络提炼为定价
5. 智能的预测带来的经济收益很大