Forecasting Models for Inventory and Demand Planning
Choosing the right forecasting models is one of the most overlooked decisions in setting up demand and inventory planning. The best inventory forecasting software doesn’t rely on a single technique – it applies the model that fits your data, from simple moving averages for stable, low-volume items to machine learning models for complex, high-variability demand.
Simple Time-Series Models
These are the starting point for most demand forecasting work, and they remain useful precisely because they’re easy to understand and quick to apply:
Simple Moving Average (SMA)
SMA predicts the next value as the average of the last n observations. It’s the most basic time-series forecasting technique, and a reasonable baseline before trying anything more complex.
Weighted Moving Average (WMA)
WMA assigns linear weights to each of the past n observations, giving more recent data more influence over the forecast than older data.
Exponential Moving Average (EMA)
EMA predicts the next observation as the exponentially weighted mean of past observations, so recent values matter more without a hard cutoff like WMA uses.
Seasonal Forecasting Models
Holt-Winters (Triple Exponential Smoothing)
Holt-Winters is a time-series method built for seasonal data. It accounts for level, trend, and seasonality using three smoothing parameters, and comes in two variations – multiplicative and additive – depending on whether seasonal swings grow with the trend or stay constant.
Statistical Forecasting Models
ARIMA
AutoRegressive Integrated Moving Average (ARIMA) is the generalized version of ARMA models. For a stationary time series, its equation is linear, with predictors made up of lags of the dependent variable and/or lags of forecast errors. It’s a strong choice when you have a longer, cleaner history and need more precision than a moving average gives you. For a deeper technical treatment, Rob Hyndman’s ARIMA models chapter is the standard reference.
GARCH and ARCH
Generalised Autoregressive Conditional Heteroskedasticity (GARCH) is used by financial professionals to model volatile markets in three steps: estimate the best-fitting autoregressive model, compute autocorrelations in the error term, and then test for significance. Autoregressive Conditional Heteroskedasticity (ARCH) is the simpler predecessor, used for modelling time series with time-varying volatility, most often where there are short bursts of increased variation.
Machine Learning Forecasting Models
Artificial Neural Networks (ANN)
ANNs consist of interconnected nodes modelled loosely on neurons – each node takes an input, operates on it, and passes a result to other nodes. A feedforward ANN has a one-directional flow of information with no feedback loops and is commonly used in pattern recognition. A feedback ANN, by contrast, includes feedback loops that let the network refine its output over multiple passes.
LSTM Networks
Long Short-Term Memory (LSTM) networks are a type of recurrent neural network capable of remembering information over long periods, specifically designed to avoid the long-term dependency problem that trips up simpler recurrent models. LSTM is behind many of the more advanced results in modern deep learning-based forecasting.
Regression-Based Forecasting
Regression measures the strength of the relationship between a dependent variable and one or more independent variables. Simple linear regression uses one independent variable to predict the dependent variable; multiple linear regression uses more than one, which is usually the more realistic case for real demand data influenced by several factors at once.
Choosing the Right Model for Your Business
No single model wins across every situation – the right choice depends on how much history you have, whether your data is seasonal, and how much volatility you need to account for. A stable, slow-moving SKU might do fine on a simple moving average; a seasonal product line usually needs Holt-Winters or ARIMA; genuinely complex, high-variability demand is where machine learning models like ANN or LSTM start to earn their added complexity. This is exactly the kind of model-selection work SAIT’s data science team handles as part of building out demand forecasting software for a specific business, rather than forcing every dataset through the same technique.
Want help matching the right forecasting model to your data? Call us to discuss our data science competencies in TensorFlow and Python/R programming for demand forecasting.