AWS aws-supply-chain documentation change
Summary
Added 'Demand History Requirement' column to forecast algorithm comparison table, specifying minimum historical data needed for each model (2x-4x forecast horizon). Updated footer text from 'First time configuration of Demand Plan' to 'Demand Pattern and Recommendation Report Access'.
Security assessment
The changes add operational requirements for historical data length but contain no evidence of addressing security vulnerabilities or security features. The updates focus on model performance prerequisites rather than security controls, access management, or vulnerability mitigation.
Diff
diff --git a/aws-supply-chain/latest/userguide/forecast-algorithims.md b/aws-supply-chain/latest/userguide/forecast-algorithims.md index 0486b48b3..7e309b487 100644 --- a//aws-supply-chain/latest/userguide/forecast-algorithims.md +++ b//aws-supply-chain/latest/userguide/forecast-algorithims.md @@ -11,27 +11,27 @@ The following table lists the 25 built-in forecast models offered by AWS Supply -Type | Forecast Ensembler/Algorithm | Model(s) in Ensemble | Automated hyper Parameter Tuning (Yes/No) | Default Parameters | Metric Optimized | Scenario(s) the model is best suited for | Supports Related Times as Forecast Inputt - Yes/No? ----|---|---|---|---|---|---|--- -Forecast Model(s) Ensembler | AutoGluon Best Quality (MAPE) | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | MAPE (Mean Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series -Forecast Model(s) Ensembler | AutoGluon Best Quality (WAPE) | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | WAPE (Weighted Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series -Forecast Model(s) Ensembler | AutoGluon Best Quality (MASE) | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | MASE (Mean Absolute Scaled Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series -Forecast Model(s) Ensembler | AutoGluon Best Quality (RMSE) | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | RMSE (Root Mean Squared Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series -Forecast Model(s) Ensembler | AutoGluon Best Quality (WCD) | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | WCD (Weighted Cumulative Deviation) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series -Forecast Model(s) Ensembler | AutoGluon StatEnsemble (MAPE) | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | MAPE (Mean Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | No -Forecast Model(s) Ensembler | AutoGluon StatEnsemble (WAPE) | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | WAPE (Weighted Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | No -Forecast Model(s) Ensembler | AutoGluon StatEnsemble (MASE) | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | MASE (Mean Absolute Scaled Error) | Automated Ensemble without need for manual model assignment/selection. | No -Forecast Model(s) Ensembler | AutoGluon StatEnsemble (RMSE) | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | RMSE (Root Mean Squared Error) | Automated Ensemble without need for manual model assignment/selection. | No -Forecast Model(s) Ensembler | AutoGluon StatEnsemble (WCD) | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | WCD (Weighted Cumulative Deviation | Automated Ensemble without need for manual model assignment/selection. | No -Forecast Model(s) Ensembler | AWS Supply Chain AutoML | Ensemble of all in [Amazon Forecast AutoML](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-choosing-recipes.html). | Not Applicable | AutoML default settings | WQL (Weighted Quantile Loss) for P10, P50, P90 | Automated Ensemble without need for manual model assignment/selection. | Depends on Selected Models by Ensembler. -Forecast Algorithm | CNN-QR | CNN-QR (Convolutional Neural Network - Quantile Regression) is a machine learning algorithm for time series forecasting using causal convolutional neural networks (CNNs). | Not Applicable | [CNN-based parameters](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-algo-cnnqr.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for large datasets containing hundreds of time series. | Yes, Past and Future Related Time Series -Forecast Algorithm | DeepAR+ | DeepAR+ is a machine learning algorithm for time series forecasting using recurrent neural networks (RNNs). | Not Applicable | [DeepAR default settings](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-deeparplus.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for large datasets containing hundreds of time series. | Yes, Past and Future Related Time Series -Forecast Algorithm | LightGBM | Light Gradient-Boosting Machine (LGBM) is a tabular machine learning model that uses historical demand data from past seasons. | Not Applicable | LightGBM default parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for datasets where different items share similar demand trends. Less effective on datasets with diverse item characteristics and demand patterns. | No -Forecast Algorithm | Prophet | Prophet is a time series forecasting algorithm based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality. | Not Applicable | [Default Prophet settings](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-prophet.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for time series that have strong seasonal effects and several seasons of historical data. | Yes, Past and Future Related Time Series -Forecast Algorithm | Triple Exponential Smoothing | Exponential Smoothing (ETS) is a statistical model for time series forecasting. | Not Applicable | Default ETS parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for datasets with seasonality patterns, computing weighted averages of past observations with exponentially decreasing weights. ETS is most effective for time series with fewer than 100 items. | No -Forecast Algorithm | Auto Complex Exponential Smoothing (AutoCES) | Auto Complex Exponential Smoothing is an advanced variant of exponential smoothing, automatically adjusts smoothing parameters, offering accurate forecasts for time series with intricate seasonal structures. | Not Applicable | [Default AutoCES settings](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.AutoCESModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for complex seasonal patterns in time series data, including multiple seasonality or irregular cycles. | No -Forecast Algorithm | ARIMA | ARIMA (Auto-Regressive Integrated Moving Average) is a statistical model for time series forecasting. It combines autoregressive, moving average, and differencing components to model trends. | Not Applicable | [ARIMA default parameters](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-arima.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for datasets without strong seasonal effects. | No -Forecast Algorithm | Seasonal ARIMA | SARIMA (Seasonal Auto-Regressive Integrated Moving Average) is an extension of ARIMA that includes seasonal components, It models both non-seasonal and seasonal trends, ensuring accurate predictions for datasets with multiple seasons of historical data. | Not Applicable | Seasonal ARIMA default parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for time series with strong seasonal patterns. | No -Forecast Algorithm | Theta | The Theta model is a time series forecasting method that combines exponential smoothing with a decomposition approach to handle trend, seasonality, and noise. It uses a linear trend model and non-linear smoothing components to capture both short-term and long-term patterns, often outperforming traditional methods. | Not Applicable | [Theta method default settings](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.ThetaModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for intermittent demand forecasting. | No -Forecast Algorithm | Aggregate-Disaggregate Intermittent Demand Approach (ADIDA) | ADIDAaggregates data at a higher level to capture broader patterns, then disaggregates it for accurate forecasts improves accuracy by reducing noise. | Not Applicable | [ADIDA default parameters](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.ADIDAModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for products with low or irregular demand, intermittent demand. | No -Forecast Algorithm | Croston | The Croston method is designed for intermittent demand forecasting. It separates demand into two components the size of non-zero demands and the intervals between them. These components are independently forecasted and combined. | Not Applicable | [Croston default settings](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.CrostonModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for intermittent demand forecasting. | No -Forecast Algorithm | Intermittent Multiple Aggregation Prediction Algorithm (IMAPA) | IMAPA is a forecasting method for intermittent demand data, where demand is irregular with many zero values. It aggregates data at multiple levels to capture different demand patterns, offering more robust predictions for datasets with highly irregular demand compared to methods like Croston. | Not Applicable | [IMAPA default parameters](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.IMAPAModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for improving accuracy for intermittent demand patterns (compared to traditional methods like exponential smoothing). | No -Forecast Algorithm | Moving Average | The Moving Average model forecasts by averaging past data points over a fixed window. | Not Applicable | Moving Average default parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for short-term forecasts, especially in sparse data scenarios. This method performs well on time series with simple trends, providing quick, easy predictions without requiring complex modeling. | No -Forecast Algorithm | Non Parametric Time Series (NPTS) | NPTS is a baseline forecasting method for sparse or intermittent time series data. It includes variants such as Standard NPTS and Seasonal NPTS. | Not Applicable | [NPTS default parameters](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-npts.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for robust predictions for irregular time series by handling missing data and seasonal effects. It is scalable and effective for irregular demand data. | No +Type | Forecast Ensembler/Algorithm | Demand History Requirement | Model(s) in Ensemble | Automated hyper Parameter Tuning (Yes/No) | Default Parameters | Metric Optimized | Scenario(s) the model is best suited for | Supports Related Times as Forecast Inputt - Yes/No? +---|---|---|---|---|---|---|---|--- +Forecast Model(s) Ensembler | AutoGluon Best Quality (MAPE) | At least 2 times the forecast horizon | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | MAPE (Mean Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series +Forecast Model(s) Ensembler | AutoGluon Best Quality (WAPE) | At least 2 times the forecast horizon | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | WAPE (Weighted Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series +Forecast Model(s) Ensembler | AutoGluon Best Quality (MASE) | At least 2 times the forecast horizon | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | MASE (Mean Absolute Scaled Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series +Forecast Model(s) Ensembler | AutoGluon Best Quality (RMSE) | At least 2 times the forecast horizon | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | RMSE (Root Mean Squared Error) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series +Forecast Model(s) Ensembler | AutoGluon Best Quality (WCD) | At least 2 times the forecast horizon | Ensemble of baseline, statistical , ML/Deep learning models in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library. | Yes | AutoGluon best_quality preset | WCD (Weighted Cumulative Deviation) | Automated Ensemble without need for manual model assignment/selection. | Yes, Past and Future Related Time Series +Forecast Model(s) Ensembler | AutoGluon StatEnsemble (MAPE) | At least 2 times the forecast horizon | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | MAPE (Mean Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | No +Forecast Model(s) Ensembler | AutoGluon StatEnsemble (WAPE) | At least 2 times the forecast horizon | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | WAPE (Weighted Absolute Percentage Error) | Automated Ensemble without need for manual model assignment/selection. | No +Forecast Model(s) Ensembler | AutoGluon StatEnsemble (MASE) | At least 2 times the forecast horizon | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | MASE (Mean Absolute Scaled Error) | Automated Ensemble without need for manual model assignment/selection. | No +Forecast Model(s) Ensembler | AutoGluon StatEnsemble (RMSE) | At least 2 times the forecast horizon | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | RMSE (Root Mean Squared Error) | Automated Ensemble without need for manual model assignment/selection. | No +Forecast Model(s) Ensembler | AutoGluon StatEnsemble (WCD) | At least 2 times the forecast horizon | Ensemble of all statistical models(only) in the [AutoGluon](https://auto.gluon.ai/stable/index.html) model library eto produce forecasts. | Yes | AutoGluon all Supported Stats Model | WCD (Weighted Cumulative Deviation | Automated Ensemble without need for manual model assignment/selection. | No +Forecast Model(s) Ensembler | AWS Supply Chain AutoML | At least 2 times the forecast horizon | Ensemble of all in [Amazon Forecast AutoML](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-choosing-recipes.html). | Not Applicable | AutoML default settings | WQL (Weighted Quantile Loss) for P10, P50, P90 | Automated Ensemble without need for manual model assignment/selection. | Depends on Selected Models by Ensembler. +Forecast Algorithm | CNN-QR | At least 4 times the forecast horizon | CNN-QR (Convolutional Neural Network - Quantile Regression) is a machine learning algorithm for time series forecasting using causal convolutional neural networks (CNNs). | Not Applicable | [CNN-based parameters](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-algo-cnnqr.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for large datasets containing hundreds of time series. | Yes, Past and Future Related Time Series +Forecast Algorithm | DeepAR+ | At least 4 times the forecast horizon | DeepAR+ is a machine learning algorithm for time series forecasting using recurrent neural networks (RNNs). | Not Applicable | [DeepAR default settings](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-deeparplus.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for large datasets containing hundreds of time series. | Yes, Past and Future Related Time Series +Forecast Algorithm | LightGBM | At least 2 times the forecast horizon | Light Gradient-Boosting Machine (LGBM) is a tabular machine learning model that uses historical demand data from past seasons. | Not Applicable | LightGBM default parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for datasets where different items share similar demand trends. Less effective on datasets with diverse item characteristics and demand patterns. | No +Forecast Algorithm | Prophet | At least 4 times the forecast horizon | Prophet is a time series forecasting algorithm based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality. | Not Applicable | [Default Prophet settings](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-prophet.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for time series that have strong seasonal effects and several seasons of historical data. | Yes, Past and Future Related Time Series +Forecast Algorithm | Triple Exponential Smoothing | At least 4 times the forecast horizon | Exponential Smoothing (ETS) is a statistical model for time series forecasting. | Not Applicable | Default ETS parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for datasets with seasonality patterns, computing weighted averages of past observations with exponentially decreasing weights. ETS is most effective for time series with fewer than 100 items. | No +Forecast Algorithm | Auto Complex Exponential Smoothing (AutoCES) | At least 2 times the forecast horizon | Auto Complex Exponential Smoothing is an advanced variant of exponential smoothing, automatically adjusts smoothing parameters, offering accurate forecasts for time series with intricate seasonal structures. | Not Applicable | [Default AutoCES settings](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.AutoCESModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for complex seasonal patterns in time series data, including multiple seasonality or irregular cycles. | No +Forecast Algorithm | ARIMA | At least 4 times the forecast horizon | ARIMA (Auto-Regressive Integrated Moving Average) is a statistical model for time series forecasting. It combines autoregressive, moving average, and differencing components to model trends. | Not Applicable | [ARIMA default parameters](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-arima.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for datasets without strong seasonal effects. | No +Forecast Algorithm | Seasonal ARIMA | At least 2 times the forecast horizon | SARIMA (Seasonal Auto-Regressive Integrated Moving Average) is an extension of ARIMA that includes seasonal components, It models both non-seasonal and seasonal trends, ensuring accurate predictions for datasets with multiple seasons of historical data. | Not Applicable | Seasonal ARIMA default parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for time series with strong seasonal patterns. | No +Forecast Algorithm | Theta | At least 2 times the forecast horizon | The Theta model is a time series forecasting method that combines exponential smoothing with a decomposition approach to handle trend, seasonality, and noise. It uses a linear trend model and non-linear smoothing components to capture both short-term and long-term patterns, often outperforming traditional methods. | Not Applicable | [Theta method default settings](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.ThetaModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for intermittent demand forecasting. | No +Forecast Algorithm | Aggregate-Disaggregate Intermittent Demand Approach (ADIDA) | At least 2 times the forecast horizon | ADIDAaggregates data at a higher level to capture broader patterns, then disaggregates it for accurate forecasts improves accuracy by reducing noise. | Not Applicable | [ADIDA default parameters](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.ADIDAModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for products with low or irregular demand, intermittent demand. | No +Forecast Algorithm | Croston | At least 2 times the forecast horizon | The Croston method is designed for intermittent demand forecasting. It separates demand into two components the size of non-zero demands and the intervals between them. These components are independently forecasted and combined. | Not Applicable | [Croston default settings](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.CrostonModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for intermittent demand forecasting. | No +Forecast Algorithm | Intermittent Multiple Aggregation Prediction Algorithm (IMAPA) | At least 2 times the forecast horizon | IMAPA is a forecasting method for intermittent demand data, where demand is irregular with many zero values. It aggregates data at multiple levels to capture different demand patterns, offering more robust predictions for datasets with highly irregular demand compared to methods like Croston. | Not Applicable | [IMAPA default parameters](https://auto.gluon.ai/dev/tutorials/timeseries/forecasting-model-zoo.html#autogluon.timeseries.models.IMAPAModel) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for improving accuracy for intermittent demand patterns (compared to traditional methods like exponential smoothing). | No +Forecast Algorithm | Moving Average | At least 2 times the forecast horizon | The Moving Average model forecasts by averaging past data points over a fixed window. | Not Applicable | Moving Average default parameters | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for short-term forecasts, especially in sparse data scenarios. This method performs well on time series with simple trends, providing quick, easy predictions without requiring complex modeling. | No +Forecast Algorithm | Non Parametric Time Series (NPTS) | At least 4 times the forecast horizon | NPTS is a baseline forecasting method for sparse or intermittent time series data. It includes variants such as Standard NPTS and Seasonal NPTS. | Not Applicable | [NPTS default parameters](https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-npts.html) | WQL (Weighted Quantile Loss) for P10, P50, P90 | Best suited for robust predictions for irregular time series by handling missing data and seasonal effects. It is scalable and effective for irregular demand data. | No @@ -56 +56 @@ To use the Amazon Web Services Documentation, Javascript must be enabled. Please -First time configuration of Demand Plan +Demand Pattern and Recommendation Report Access