Skip to main content

The Model Arena

Author: Drew Shea, Created: 2024-05-23

There are many machine learning and statistical forecasting models for a reason. Some forecasting models are great at capturing multiple seasonal fluctuations, while other models excel at predicting underlying trends. Even within the same business, data patterns can vary wildly per region, per product, per service, etc. Why would you want a one-size-fits-all approach to your forecasting process across a variety of different line items (products, regions, services, plants, etc. intersections)?

That is why SensibleAI Forecast has the Model Arena. The SensibleAI Forecast Model Arena is a proprietary modeling technique that is purpose-built to squeeze every ounce of forecast accuracy out of the line items to be forecasted in your business planning processes.

At its core, the Model Arena is a competitive model training environment where multiple machine learning and statistical models compete against each other for a chance to be elected as the “champion” or the best performer for each of the respective line items to be forecasted. This process is grounded in a rigorous training strategy that ensures models are not only accurate but robust across various time periods to ensure that the models do not overfit the training data as business conditions change. By the Model Arena leveraging a diverse set of models, SensibleAI Forecast can choose the right model for the job based on the unique data patterns that exist for each line item to be forecasted to make sure you are not leaving any accuracy on the table.

Please see the table below for the list of forecast models that exist within the Model Arena.

Model Name

Category

Description

Interpretability

Supports Multiseries

XGBoost

ML

Extreme Gradient Boosting; known for its performance and speed, excellent for handling large datasets and complex patterns.

Feature Impact

Prediction Explanations

Catboost

ML

Gradient boosting with categorical features support; excels in handling categorical data and preventing overfitting.

Feature Impact

Prediction Explanations

LGBM

ML

LightGBM, a gradient boosting framework; highly efficient and fast, suitable for large datasets with many features.

Feature Impact

Prediction Explanations

Cubist

ML

Rule-based regression model; effective for complex relationships and handling both numeric and categorical predictors.

Feature Impact

Prediction Explanations

Random Forest

ML

Ensemble of decision trees; robust and effective for capturing non-linear relationships and interactions.

Feature Impact

Prediction Explanations

Elastic Net

ML

Combines L1 and L2 regularization; useful for feature selection and handling multicollinearity in large datasets.

Feature Impact

Prediction Explanations

Poly Elastic Net

ML

Combines polynomial features with Elastic Net; useful for capturing non-linear relationships.

Feature Impact

Prediction Explanations

SVR

ML

Support Vector Regression; handles non-linear relationships well, effective for small to medium-sized datasets.

Feature Impact

Prediction Explanations

Tweedie GLM

ML

Generalized Linear Model with Tweedie distribution; good for modeling data with both continuous and discrete components.

Feature Impact

Prediction Explanations

Arima

Statistical

Autoregressive Integrated Moving Average; powerful for series with trends and seasonality when properly tuned.

Feature Impact

Prediction Explanations

Croston

Statistical

Specifically designed for intermittent demand forecasting; good for sparse data with periods of zero demand.

Feature Impact

Prediction Explanations

ETS Simple

Statistical

Exponential Smoothing; best for capturing level and trend components in time series.

Feature Impact

Prediction Explanations

Exponential Smoothing

Statistical

Weighs past observations with exponentially decreasing weights; useful for smoothing and forecasting.

Feature Impact

Prediction Explanations

Fourier

Statistical

Decomposes time series into sinusoidal components; effective for capturing periodic patterns.

Feature Impact

Prediction Explanations

Holt Linear

Statistical

A specific case of Holt’s method with linear trend; best for series with linear trends.

Feature Impact

Prediction Explanations

Holt Winter Additive

Statistical

Adds seasonal components to Holt's method; good for series with additive seasonality.

Feature Impact

Prediction Explanations

Holt Winter Multiplicative

Statistical

Adds seasonal components multiplicatively; best for series with multiplicative seasonality.

Feature Impact

Prediction Explanations

Seasonal Additive

Statistical

Models seasonality additively; suitable for series with additive seasonal patterns.

Feature Impact

Prediction Explanations

Seasonal Multiplicative

Statistical

Models seasonality multiplicatively; effective for series with multiplicative seasonal patterns.

Feature Impact

Prediction Explanations

Simple Exponential Smoothing

Statistical

Single-parameter smoothing; best for level data without trend or seasonality.

Feature Impact

Prediction Explanations

Simple Moving Average

Statistical

Averages a fixed number of past observations; useful for smoothing out short-term fluctuations.

Feature Impact

Prediction Explanations

Theta

Statistical

Combines decomposed components of time series; known for its simplicity and effectiveness.

Feature Impact

Prediction Explanations

Mean

Baseline

Averages all past values to forecast future values; best for stationary series without trend or seasonality.

Feature Impact

Prediction Explanations

Shift

Baseline

Uses lagged values as predictors; good for capturing autocorrelation in the data.

Feature Impact

Prediction Explanations

Last Value Naïve

Baseline

Uses the last observed value as the forecast; simple and effective for random walks or very short-term forecasts.

Feature Impact

Prediction Explanations

Was this page helpful?