In this project, we will implement a very aged prediction problem from the financial field.From a series of stock prices, including daily open, high, low, and close prices, decide our daily action and make our best profit for the future trading.
XGBoost works as Newton-Raphson in function space unlike gradient boosting that works as gradient descent in function space, a second order Taylor approximation is used in the loss function to make the connection to Newton Raphson method. 詳細內容可以再參照 (Tianqi Chen&Carlos Guestrin,2016)
我們這次使用 XGBoost 演算法， 將股票前20天的股市 opening price 當作 features， feed 給 XGBoost 演算法當作已知的 feature (X)， 再嘗試 predict 出 表示當日與明日股票的升降狀況 y label。
- 先預測股票升降
- 股價升：得到的
predict y == 1 - 股價降：得到的
predict y == 0
- 股價升：得到的
- 再根據升降的預測結果和當前所持有的股票數判斷
明天股價變化 目前持有股票數 行動 升 1 持平 降 1 賣出 升 0 買入 降 0 賣出 升 -1 買入 降 -1 持平
| Open | Prediction | Action | Shares | Balance |
|---|---|---|---|---|
| 154.4 | bull | 1 | 1 | -154.4 |
| 155.96 | bear | -1 | 0 | 1.5600000000000023 |
| 156.45 | bear | -1 | -1 | 158.01 |
| 154.1 | bull | 1 | 0 | 3.9099999999999966 |
| 153.59 | bull | 1 | 1 | -149.68 |
| 154.81 | bull | 0 | 1 | -149.68 |
| 155.46 | bull | 0 | 1 | -149.68 |
| 156.74 | bear | -1 | 0 | 7.060000000000002 |
| 156.6 | bull | 1 | 1 | -149.54 |
| 154.6 | bull | 0 | 1 | -149.54 |
| 153.61 | bull | 0 | 1 | -149.54 |
| 153.59 | bull | 0 | 1 | -149.54 |
| 154.05 | bull | 0 | 1 | -149.54 |
| 153.65 | bear | -1 | 0 | 4.110000000000014 |
| 153.17 | bull | 1 | 1 | -149.05999999999997 |
| 151.82 | bull | 0 | 1 | -149.05999999999997 |
| 152.51 | bull | 0 | 1 | -149.05999999999997 |
| 152.95 | bull | 0 | 1 | -149.05999999999997 |
| 153.2 | bull | 0 | 1 | -149.05999999999997 |
| 154.17 |
Profit: 4.360000000000014
Accuracy: 0.631578947368421