Time series prediction
- 在主要feature是时间时用的比较多。如果有很多额外的feature,考虑深度模型
- Trend: 长期的趋势
- Seasonality:季节性
- Cyclicity:周期性 比如经济周期
- Irregularity:unpredictable factors。最好在数据层面就剔除掉
# Before forecasting
- Make sure the series is stationary : devoid of seasonality and trend, the mean and variance is nearly constant.
- Data analysis, decompose the four factors
- check if the 'random' component is stationary. Null hypothesis : non-stationary. Given this hypothesis, the possiblity to have the current data is the p-value.
# Models
- ARMA : Auto Regression (使用历史值回归) + Moving Average(使用历史均值回归)
- ARIMA: ARMA by 增量数据
# Correlation test
ACF:auto-correlation function. x=2, y is the correlation between x_t and x_{t-2}.
PACF: 在ACF中, x_t和x_{t-2}的correlation中间还有x_{t-1}的影响。所以有了partial ACF,把x_{t-1}的影响去掉. 具体涉及一些统计知识,需要进一步挖掘
ACF和PACF的结果指导了ARIMA的参数选择。ACF图结合significance level看。