If you like it, GAN it - Probabilistic MTS Forecast with GAN (2020, 9)
Contents
- Abstract
- Introduction
- Notation
- Methodology
- ProbCast : The Proposed MTS Forecasting model
- Deterministic to Probabilistic
- Train Pipeline
- Dataset
0. Abstract
propose ProbCast
- novel probabilistic model for MTS forecasting
- employ conditional GAN framework to train our model
1. Introduction
propose ProbCast
- probabilistic forecast model
- for MTS
- based on CGAN
In addition to ProbCast, suggest a framework…
- for transforming an existing deterministic forecaster to probabilistic one
2. Notation
MTS : \(X=\left\{X_{0}, X_{1}, \ldots, X_{T}\right\}\)
-
where \(X_{t}=\left\{x_{t, 1}, x_{t, 2}, \ldots, x_{t, f}\right\}\).
( \(f\) : number of features )
-
goal : model \(P\left(X_{t+1} \mid X_{t}, . ., X_{0}\right)\)
3. Methodology
(1) ProbCast : The Proposed MTS Forecasting model
Model : \(P\left(X_{t+1} \mid X_{t}, . ., X_{0}\right)\)
- condition : \(\left\{X_{t}, . ., X_{0}\right\}\)
Value function
- \(\begin{aligned} \min _{P C} \max _{D} V(D, P C)=& \mathbb{E}_{X_{t+1} \sim P_{\text {data }}\left(X_{t+1}\right)}\left[\log D\left(X_{t+1} \mid X_{t}, . ., X_{0}\right)\right]+\\ & \mathbb{E}_{z \sim P_{z}(z)}\left[\log \left(1-D\left(P C\left(z \mid X_{t}, . ., X_{0}\right)\right)\right)\right] \end{aligned}\).
(2) Deterministic to Probabilistic
in MTS, need to figure out “dependencies between features”
propose a new framework for building “probabilistic forecaster”, based on deterministic forecaster using GAN
(3) Train Pipeline
-
Step 1) build an accurate point forecast model
- Step 2) integrate the noise vector \(z\)into the deterministic model architecture
- Step 3) train this model
4. Dataset
electricity & exchange rate datasets
