ulandz / a-deep-rl-framework-for-futures-trading

A-Deep-RL-Framework-for-Index-futures-Trading

Data

• 台指期
  • Source:
    • DQ2: 1987/1/7 ~ 2021/3/15
    • 期交所: 2021/3/15 ~
  • 交易時間: 前日夜盤+今日日盤
  • 其他說明:
    • DQ2成交量: 不含價差交易

feature

• log return
  • $\ln\frac{P_t}{P_{t-1}}$, $P_t$為資產在t時刻的價格
• 基差(basis)
  • def: 現貨價格 - 期貨價格
  • $\frac{現貨-期貨}{現貨}\times10$
    • 乘10目的是放大原始值域([-0.03, 0.07]->[-0.3, 0.7])
• 成交量/OI
  • percentile: data[:t], 統計t時刻值在t時刻前所有資料的百分位數
    • 刻劃破新高
  • deg_change: $\arctan(V_t-V_{t-1}, 100)$
    • 刻劃變化量, 100為歷史成交量下限
• K棒
  • 實體: $|o-c|$
  • 上影線: $high-\max(o, c)$
  • 下影線: $\min(o,c)-low$
  • 除range(h-l), 以分形角度看K棒
• OHLC
  • Drift‐Independent Volatility Estimation Based on High, Low, Open, and Close Prices
  • $o_t = \ln(O_t)-\ln(C_{t-1})$
  • $u_t = \ln(H_t)-\ln(O_t)$
  • $d_t = \ln(L_t)-\ln(O_t)$
  • $c_t = \ln(C_t)-\ln(O_t)$
• Hurst exponent
  • Hurst Exponent and Trading Signals Derived from Market Time Series
  • $p(k) = Ck^{-\alpha}$
    • p(k) is an autocorrelation function
    • C is constant
    • k is number of lags
    • $\alpha$ is the decay parameter
      • $\alpha$ ranges between [0,2]
      • $0<\alpha<1$: persistence
      • $\alpha=1$: random walk
      • $1<\alpha<2$: mean reversion
  • $H=1-\frac{\alpha}{2}$
    • H=0.5: random walk
    • H<0.5: mean reversion
    • H>0.5: momentum
  • 2Q/4Q

ref

fractal

• the first 100 pages of the book,"The Science of Fractal Images" edited by Heinz-Otto Peitgen and Dietmar Saupe

FinRL

https://towardsdatascience.com/deep-reinforcement-learning-for-automated-stock-trading-f1dad0126a02

todo

• use Dask df to improve speed

note

pip install numpy==1.19.5