接上篇：Forecasting Big Time Series: Theory and Practice（Part I）
文章目錄
- Forecasting with Neural Networks – Timeline
- Deep Learning for Forecasting: Outline
- Feed-Forward Neural Networks (Multi-layer Perceptron (MLPs))
- Basic Model Structures
- Sequence-to-Sequence / Many-to-Many Structure (Discriminative Model)
- Training Neural Networks
- Loss Functions
- Recurrent Neural Networks
- Recap: State-Space Models for Forecasting
- From Latent State (Exponential Smoothing) to Recurrent NN
- Central idea: Exponential Smoothing
- Long Short-Term Memory (LSTM)
- Basic Model Structures
- Canonical RNN Structure: DeepAR
- Sequence to Sequence (Seq2Seq) Structure: Many-to-Many
- Seq2Seq: RNN-MLP
- Seq2Seq: RNN-RNN
- Seq2Seq: Causal CNN-RNN
- Seq2Seq: Training
- Seq2Seq: Gated Recurrent Units (GRU) [Suilin, 2017]
- Comparison: Canonical (One-to-One) vs. Seq2Seq (Many-to-Many)
- Probabilistic Forecasts from Neural Nets
- How to Get Probabilistic Forecasts
- Input/Output Transformations
- Input/Output Transformations: Box-Cox Transform
- Input/Output Transformations: Probabilty Integral Transform
- Input/Output Transformations: Binning
- Distribution Representations
- Spline Quantile Function, [Gasthaus et al., 2019]
- GANs for Forecasting
- Convolutional Neural Networks
- Dilated Causal Convolution and WaveNet [Van Den Oord et al., 2016]
- What about attention?
- Transformer for Forecasting
- N-BEATS: [Oreshkin et al., 2019]
- Modern methods struggle with strategic forecasting problems
- Deep Probabilistic Models
Forecasting with Neural Networks – Timeline
Deep Learning for Forecasting: Outline
Feed-Forward Neural Networks (Multi-layer Perceptron (MLPs))
Basic Model Structures
Sequence-to-Sequence / Many-to-Many Structure (Discriminative Model)
Training Neural Networks
Loss Functions
Recurrent Neural Networks
Recap: State-Space Models for Forecasting
From Latent State (Exponential Smoothing) to Recurrent NN
Central idea: Exponential Smoothing
Long Short-Term Memory (LSTM)
Basic Model Structures
Canonical RNN Structure: DeepAR
Sequence to Sequence (Seq2Seq) Structure: Many-to-Many
Seq2Seq: RNN-MLP
Seq2Seq: RNN-RNN
Seq2Seq: Causal CNN-RNN
Seq2Seq: Training
Seq2Seq: Gated Recurrent Units (GRU) [Suilin, 2017]
Comparison: Canonical (One-to-One) vs. Seq2Seq (Many-to-Many)
Probabilistic Forecasts from Neural Nets
How to Get Probabilistic Forecasts
Input/Output Transformations
Input/Output Transformations: Box-Cox Transform
Input/Output Transformations: Probabilty Integral Transform
Input/Output Transformations: Binning
Distribution Representations
Spline Quantile Function, [Gasthaus et al., 2019]
GANs for Forecasting
Convolutional Neural Networks
Dilated Causal Convolution and WaveNet [Van Den Oord et al., 2016]
What about attention?
Transformer for Forecasting
N-BEATS: [Oreshkin et al., 2019]
Modern methods struggle with strategic forecasting problems
Predict overall Amazon retail demand years into the future. Not enough data may be available for training, assumptions on long-term behaviour should be handled properly. Use a classical, local model
預測未來亞馬遜的總體零售需求。可能沒有足夠的數據用於培訓，應正確處理有關長期行爲的假設。使用經典的本地模型
Predict the demand for a each product available at Amazon Time series are irregular, only combined to they have enough history and exhibit clear patterns.
預測對亞馬遜時間序列中可用的每種產品的需求都是不規則的，只有結合起來，它們才具有足夠的歷史並顯示出清晰的模式。
Deep Probabilistic Models
Finding the right balance: data vs model driven
找到合適的平衡：數據與模型驅動
Simple Exponential Smoothing
General Exponential Smoothing
Linear State Space Model（SSM）
Deep State Space Model in a Nutshell
Deep State - Training
Explore Structure: Local vs. Global
Deep Factor Models: Local & Global
Summary of Deep State and Deep Factor
