Forecasts
Demand forecasting is the process of estimating the future demand for a stock item. The output of this process is a forecast, which is a data-driven estimate of the quantity of a product that consumers will demand in a forthcoming period. The forecasts calculated by Inventory Optimization are based on the historical data outlined in the previous section. An example of a demand forecast is seen below.
Horizon
A forecast has a horizon, referring to the future time period over which the forecast is made. Forecast horizons can be short-term (a few weeks), medium-term (a few months), or long-term (up to a year). Setting the forecast horizon depends on how far ahead the planner needs insight when placing a purchase order for a stock item.
Granularity
The granularity of a forecast determines whether a forecast is on a daily, weekly, or monthly level, affecting its level of detail. For example, if the granularity is weekly, and the horizon is 4, the forecast is made for the next 4 weeks, with one estimate per week.
Forecasting with monthly granularity is often easier than daily or weekly forecasting. This is because monthly forecasts smooth out the daily and weekly fluctuations, reducing the impact of random noise and highlighting longer-term trends more clearly. With fewer data points to analyze, it's simpler to identify patterns and make predictions. We therefore recommend partners to implement monthly demand forecasts by default, unless the planner really needs more detail.
If a planner needs more detail, she can increase the granularity, i.e. set it to weekly or daily. We advice partners to inform users that increasing the granularity can result in less accurate forecasts, and that the granularity should be kept as low as possible while taking their needs into account.
Data Requirements
Inventory Optimization will return a demand forecast even with just one historical data sample. This is to provide a consistent service for all stock items. However, forecasts based on little data will not provide a lot of insight to the planner. Another factor is the quality of the data. 10 years of noisy and erratic data will unfortunately not result in good forecasts.
Generally, we say that the forecasting models start picking up patterns in the data at around 6 months of data. After 1 year, the forecasts can pick up on season, and this will get stronger with additional years of data. Note that planners are most interested in the most recent year of historical data, maybe the year before that as well. As a result, there is also a cap on how much data is really needed. Inventory Optimization weights the most recent years highest, so providing more than 4-5 years of data is often not necessary.
Expected Value and Upper/Lower Bounds
A forecast consists of multiple values, one per time step in the horizon. The forecast contains one expected value for each time step. An expected value is the central estimate of the demand in time step, serving as the most likely outcome based on the historical data. Accompanying the expected value, are the upper and lower bounds. This interval reflects the range within which the actual outcome is expected to fall, providing a measure of uncertainty around the forecasted value.
Decomposition
A forecast value is shaped by three key elements: season, trend, and noise. The season reflects the annual fluctuations in demand, characterized by periodic highs and lows, exemplified by increased swimsuit sales in the summer. The trend shows the overarching direction of demand over time, indicating whether interest in a product is rising, declining, or remaining constant. The noise represents the irregular and unpredictable variations in demand, arising from spontaneous and hard-to-foresee events. Together, these values add up to the forecast value.
The decomposition can be used to explain the forecast. For some products there may be no indication of seasonality. In these cases the seasonal component values will all be 0. Breaking down each forecast value can help planners gain extra insight.