ML: In Machine Learning Algorithms—Introduction to factor model (multivariate), time series model/time series model (mainly univariate) algorithm comparison, overview of factor model with time series, detailed strategy of case application
Table of contents
Introduction to Factor Models and Timing/Time Series Modeling Algorithms
2. Overview of common factor models with time series
Introduction to Factor Models and Timing/Time Series Modeling Algorithms
1. Overview of factor model (multivariate) and time series model/time series model (mainly univariate) algorithms
| Factor Model (Multivariate) |
Time series model (mainly univariate) |
|
| Introduction |
The factor model starts from the linear relationship between variables, finds common factors, and then decomposes the variables into linear combinations of several factors . In practical applications, relationships also include non-linearities. Mainly machine learning algorithms, such as LiR, RF, LightGBM and other algorithms; |
The time series model starts from the perspective of time and establishes a forecast model for the future through the analysis and fitting of historical data . Mainly traditional statistical models, such as ARIMA model; |
| features |
Utilize a variety of data that has relevance and influence on the target variable, including historical past data. |
Mainly use historical past data to model and predict; |
| data request |
Usually contains more variables : Factor models require linear relationships between variables . |
Only for one time series variable : The time series model requires the data to meet conditions such as stationarity . |
| model accuracy |
In most cases, the effect is ok, but when the assumptions of the data are not met, the prediction results may be poor ; |
The prediction accuracy is relatively high ; |
| decision guidance |
instructive |
Forecast only, no guidance |
| experience |
In practice, the factor model and the time series model are often used in combination to realize the prediction of various tasks and provide guidance; but the prediction of the time series model will take the driving factor of the factor model Explanatory power is diluted. |
|
| Three Combination Technologies |
Integrated Models - Factor-Time Series Mixed Models - Modeling Separately Forecast Weighting : Factor models and time series models are modeled separately, then forecasts are combined. In this method, the factor model and the time series model are modeled independently, and then the prediction results of the two models are weighted and averaged or other combination methods are used to obtain the final prediction result. The advantage of this method is that the two models play their respective strengths and can improve the prediction accuracy. The disadvantage is that modeling and optimization need to be carried out on the two models separately, which takes more time and effort. A single model based on feature fusion : It can be considered to fuse the influencing features extracted by the factor model with the time series data as the input features of the time series model, so as to improve the predictive ability of the time series model. It is also possible to use the output of the time series model as one of the input factors of the factor model. Adopt a single dynamic factor model : The dynamic factor model combines the characteristics of the factor model and the time series model, considers the relationship between multiple variables and the factors of the time series, and can dynamically update the factors and weights. The advantage of this method is that it solves the shortcomings of static factor models and time series models to a certain extent, but it needs to deal with the dynamics of factors. Such as BDFM, DFS-MS, DFM-MFD, TV-DFM, etc. |
|
2. Overview of common factor models with time series
| dynamic factor model |
Introduction |
core principle |
| BDFM |
Bayesian Dynamic Factor Model (BDFM): BDFM is a dynamic factor model based on Bayesian statistics , which can model the dynamics of factors and estimate the uncertainty of model parameters. |
Dynamic factor models based on Bayesian statistics. By modeling the dynamics of factors, it can model factors in time and estimate the uncertainty of model parameters through Bayesian inference methods. |
| DFS-MS |
Dynamic Factor Model with Markov Switching (DFS-MS): DFS-MS is a dynamic factor model based on Markov switching , which can capture the dynamic relationship between factors and time series, and can deal with nonlinear changes. |
Dynamic factor models based on Markov switching. It can capture the dynamic relationship between factors and time series, and can cope with nonlinear changes. The model assumes that the state of the system switches in the form of a Markov chain in different time periods, and each state corresponds to a different dynamic relationship. |
| DFM-MFD |
Dynamic Factor Model with Mixed Frequency Data (DFM-MFD): DFM-MFD is a mixed frequency dynamic factor model that can handle changes in data on different time scales and can accurately predict future changes in a longer time range. |
Dynamic factor models for mixed frequency data. It is able to handle data changes on different time scales. By combining data of different frequencies, such as high-frequency and low-frequency data, the model extracts shared dynamic factors and is able to accurately predict future changes in a longer time frame. |
| TV-DFM |
Time-varying Dynamic Factor Model (TV-DFM): TV-DFM is a time-varying dynamic factor model that can adjust factor weights according to different time periods of the data and can handle seasonal and cyclical changes. |
The core principle is a dynamic factor model based on time changes. It can adjust factor weights according to different time periods of the data to adapt to seasonal and cyclical changes in the data. By introducing time-varying factor weights, the model can more accurately capture the time evolution characteristics of the data and improve the accuracy of prediction. |
ML: In Machine Learning Algorithms—Introduction to factor model (multivariate), time series model/time series model (mainly univariate) algorithm comparison, overview of factor model with time series, detailed strategy of case application
https://yunyaniu.blog.csdn.net/article/details/130652986
Math/ML: Sequence Supervised Learning - Introduction to Time Series Datasets/Time Series Prediction Tasks (Univariate Time Series Model/Multivariate Time Series Model), Common Algorithms and Their Tools, Detailed Guide to Case Applications
https://yunyaniu.blog.csdn.net/article/details/127156732
Math/ML: Sequence Supervised Learning - Introduction to Time Series Datasets/Time Series Prediction Tasks (Univariate Time Series Model/Multivariate Time Series Model), Common Algorithms and Their Tools, Detailed Guide to Case Applications
https://yunyaniu.blog.csdn.net/article/details/127156732