Financial Time Series Data Processing for Machine Learning 笔记

论文地址：https://arxiv.org/abs/1907.03010

1.Scaling

1.MinMax

$$z=\frac{x-x_{min}}{x_{max}-x_{min}}(max-min)+min$$
max和min 是feature range。

2.Standardization

$$\begin{gathered} z=\frac{x-\mu }{\sigma } \\ \mu :mean \\ \sigma :standarddeviation \end{gathered}$$

2.Split

t r a i n = { S 0 , ⋯   , S t s − 1 } ( 1 ) v a l = { S t s , ⋯   , S t s + v s − 1 } ( 2 ) t e s t = { S t s + v s , ⋯   , S K − 1 } ( 3 ) train=\{S_0,\cdots,S_{ts-1}\}\qquad(1)\\ val=\{S_{ts},\cdots,S_{ts+vs-1}\}\qquad(2)\\ test=\{S_{ts+vs},\cdots,S_{K-1}\}\qquad(3) train={S0​,⋯,Sts−1​}(1)val={Sts​,⋯,Sts+vs−1​}(2)test={Sts+vs​,⋯,SK−1​}(3)

3.Labeling

Label	Description
N bars Up/Down	Classifier on $C_{t+n}>C_t$
N bars price change	Regressor on $C_{t+n}-C_t$
N bars log returns	Regressor on $log(\frac{C_{t+n}}{C_t})$
N bars Moving Average	Classifier on $M{A}_{t+n}>M{A}_t$
N bars trend Strength	Regressor on Trend
N bars trend Direction	Classifier on Trend
%Q after N bars	Regressor on %Q
QClass after N bars	Classifier on Qclass

  If $C_t$ is the closing price at time $t$, assume the corresponding time series slice of size $m$ ending by $C_t$ is defined by $S_t=C_{t-m+1},\cdots ,C_t$
 Then
$$\%Q_{t+1}^{t+n}=\frac{H{H}_{t+1}^{t+n}-C_t}{H{H}_{t+1}^{t+n}-L{L}_{t+1}^{t+n}}$$
 Where

1. $n$: time horizon in numbers of bars
2. $\%Q_{t+1}^{t+n}$ : %Q between $t+1$ and $t+n$
3. $H{H}_{t+1}^{t+n}$ : Highest High price between $t+1$ and $t+n$
4. $L{L}_{t+1}^{t+n}$ : Lowest Low price between $t+1$ and $t+n$

 So %Q is interpreted as following:

1. %Q = 1 when we have a perfect up move without any drawdown during the next $n$ bars
2. %Q = 0 when we have a perfect down move without any drawup during the next $n$ bars
3. %Q = 0:5 when we have an equally Up and Down move during the next $n$ bars

Class	Condition	Meaning
0	$\%Q>=0.6$	Up
1	$0.4<\%Q<0.6$	Neutral
2	$\%Q<=0.4$	Down