PlanIQ supports several algorithms that to apply to your datasets to make predictions.
This page provides information on the properties of each algorithm. This should help you decide which one to use for your forecast.
Algorithms (PlanIQ)
PlanIQ supports these algorithms:
Anaplan AutoML
Anaplan AutoML chooses the most apt algorithm for your forecast, based on the properties in your dataset. Anaplan AutoML applies all types of Amazon Forecast algorithms including hyper-parameterized variations algorithms that support that, for example DeepAR+, CNN-QR, NPTS. Anaplan AutoML automatically adjusts configuration and algorithm selection for the data collection. This algorithm doesn't include MVLR.
PlanIQ compares MASE (mean absolute scaled error) results for every individual item for every algorithm used. It then chooses the algorithm that provides the most apt MASE result for most of the items.
To summarize, Anaplan AutoML chooses the algorithm that is the best for the majority of items in the data collection. This is rather than for every item in the data collection.
For example, in a data collection with 10 items, Anaplan AutoML picks the algorithm that provides the best MASE result for the majority of items.
Based on MASE, Algorithm A provides the best forecast for three items, and Algorithm B provides the best forecast for eight items. The engine chooses Algorithm B as it provides a better performance for the majority of the items.
Amazon Ensemble
This algorithm uses the ensemble method to provide the best combination of algorithms for each item in your data collection. It also offers model-level explainability based on holidays, related data, and attributes, if provided as part of the data collection. See Understand explainability for more information.
ARIMA
ARIMA is the Autoregressive Integrated Moving Average algorithm. It's a statistical model useful for datasets that can be mapped to stationary time series. Use ARIMA when simple concepts of trend and seasonality are likely to explain most of the variance in the time series data. Since this algorithm processes each list item separately, it does not incorporate related data or attributes.
CNN-QR
This is a Convolutional Neural Network - Quantile Regression (CNN-QR) model for larger datasets that contain hundreds of time series. It's the only algorithm that accepts related data with or without forward-looking information. The algorithm also accepts attributes.
DeepAR+
This is a recurrent neural networks (RNN) model for larger data sets that contain hundreds of time series. It works with related data that has forward-looking information and attributes.
Note: CNN-QR and DeepAR+ are global algorithms. They look not only at individual time series to make predictions, but also learn from other time series from the same dataset. This is particularly useful in cases where multiple time series share similar behaviors.
Exponential Smoothing (ETS)
ETS is for time series data. ETS is useful when simple concepts of trend and seasonality are likely to explain most of the variance in the time series data. ETS is not recommended for weekly timescale data with forecast horizon of 1 year and above. Since this algorithm processes each list item separately, it's unable to take advantage of related data or attributes.
MVLR
Use a multivariate linear regression (MVLR) algorithm to generate feature-based forecasts. MVLR offers quick predictions and explainability details (contributions from the feature variables). MVLR automatically identifies and eliminates related features to improve prediction accuracy. It uses both historical data and that related data which includes forward-looking information. MVLR does not use attribute data. See Understand explainability for more information.
Prophet
This is an additive regression model. This algorithm makes use of trend, seasonal patterns and holiday impacts to generate forecasts. Prophet uses related data which includes forward-looking information. Use Prophet for scenarios that involve outliers or multiple seasonal patterns, as well as in instances where the historical time series skips some observations.
- Prophet requires at least two years of historical data, excluding the backtest period, to detect yearly seasonality patterns.
- Prophet requires at least two weeks of historical data, excluding the backtest period, to detect weekly seasonality patters. For weekly seasonality, data must be set to daily frequency.
Summary of technical differences among algorithms:
| Anaplan AutoML | Amazon Ensemble | Amazon Forecast AutoML | ARIMA | CNN-QR | DeepAR+ | ETS | MVLR | Prophet | |
| Able to use related time series | Yes | Yes | Yes | -- | Yes | Yes | -- | Yes | Yes |
| Able to use related time series without forward looking information | Yes | Yes | Yes | -- | Yes | -- | -- | -- | -- |
| Able to use item metadata | Yes | Yes | Yes | -- | Yes | Yes | -- | -- | -- |
| Suitable for what-If analysis | -- | -- | -- | -- | Yes | Yes | -- | Yes | Yes |
| Suitable for cold-start scenarios | -- | -- | Yes | -- | Yes | Yes | -- |
| -- |
| The forecast model is updated with every prediction | Depends on the optimal algorithm chosen. | -- | Underlying neural networks are updated only on retrain. To ensure full update of the model,we recommend to retrain. | Yes | -- | -- | Yes | Yes | Yes |
| The forecast model needs to retrain to incorporate new actuals into the model itself | Depends on the optimal algorithm chosen. | -- | -- | -- | Yes | Yes | -- | -- | -- |
Notes:
- Amazon Forecast AutoML will soon no longer be used within PlanIQ. We recommend you use Amazon Ensemble instead.
- Neural networks, also called simulated or artificial neural networks, are an important part of of machine learning. Neural networks are essential to the deep learning algorithms. The architecture of the human brain's neural system is the basis of the name of this group of algorithms.
Disclaimer
We update Anapedia content regularly to provide the most up-to-date instructions.