BlockRNN
BlockRNN is a type of Recurrent Neural Network (RNN) architecture used for time series forecasting. It's specifically designed to capture temporal dependencies in sequential data.
🧠 What is BlockRNN?
BlockRNN is a modular RNN forecasting model that stacks RNN layers and dense layers in blocks to predict future time series values. It is similar in structure to models like DeepAR but provides a more flexible way to configure blocks of neural layers.
🔧 Core Features
-
Sequential modeling: It uses LSTM or GRU units to capture time dependencies.
-
Probabilistic forecasting: Outputs are usually probabilistic (e.g., Gaussian or Student-T distribution), not just point estimates.
-
Flexible architecture: Can configure number of layers, hidden units, distribution output, etc.
-
Handling covariates: Supports dynamic and static features, such as holiday indicators, item categories, etc.
🔁 Architecture
Typical BlockRNN architecture includes:
-
Embedding layers for categorical inputs (like item ID, day-of-week, etc.)
-
LSTM/GRU blocks for sequential input
-
Feedforward layers post-RNN
-
Distribution projection layer that outputs parameters for a target distribution (like mean & std for Gaussian)
✅ When to Use BlockRNN
-
You have time series data with strong temporal dependencies.
-
You need probabilistic forecasts, not just point predictions.
-
You want a deep learning alternative to traditional models like ARIMA or Prophet.
📈 Pros vs. Traditional Models
| Feature | BlockRNN | ARIMA / Prophet |
|---|---|---|
| Seasonality | Learns from data | Explicitly modeled |
| Scalability | High (with GPU) | Lower for many series |
| Uncertainty | Probabilistic outputs | Some (in Prophet) |
| Feature Inputs | Supports covariates | Limited |
BlockRNN Model Parameters
| Param Name | Description | Param Value (Default) |
|---|---|---|
BLOCKRNN-TYPE |
Type of RNN block used (e.g., LSTM, GRU) in BlockRNN model. | LSTM |
HIDDEN-DIM |
Number of hidden units in each RNN layer. | 124 |
RNN-LAYERS |
Number of stacked RNN layers in the model. | 2 |
DROPOUT |
Dropout rate applied between RNN layers to reduce overfitting. | 0.0 |
EPOCHS |
Total number of epochs to train the model. | 100 |
LEARNING-RATE |
Learning rate for the optimizer used in training. | 0.001 |
BATCH-SIZE |
Number of samples per training batch. | 32 |
RANDOM-STATE |
Seed used to ensure reproducibility across runs. | 42 |
FORCE-RESET |
Whether to reset hidden states between sequences. | True |
OUTPUT-CHUNK-LENGTH |
Number of steps the model predicts in each forward pass. | 1 |