02

Forecasting Methods

The forecasting method you select is a function of multiple qualities about your item. Is demand steady, cyclical or sporadic? Are there seasonal trends? Are trends strong or limited? Is the item new? Each item being forecast has a somewhat unique history (and future), and therefore an optimal method. A method that accurately forecasts one data set might prove inaccurate for another.

Determining the optimal forecast method is a rather complex science, especially across a large product line. This may be nearly impossible using only spreadsheets. However, sophisticated forecasting software can within seconds test multiple methods for each item to determine which method will give you the most accurate results.

Moving average methods take the average of past actuals and project it forward. These methods assume that the recent past represents the future. As a result, they work best for products with relatively little change — steady demand, no seasonality, limited trends or cycles, and no significant demand shifts. Many companies apply this method because it is simple and easy to use. However, since few products actually behave in this way, it tends to be less useful than more specialized methods.

Exponential smoothing is a more advanced form of time series forecasting. Unlike moving averages, exponential smoothing methods can capture trends and recurring patterns. They accomplish this by:

Emphasizing the more recent data (as opposed to a moving average which weights all data equally), and
Smoothing out fluctuations, which are often caused by pure randomness in the data (or “noise” in the system)
Forecasters determine the forecast weights, controlling how fast or slow the model responds to demand changes in your actuals. Not all exponential smoothing methods can handle seasonality or other recurring patterns.

Exponential smoothing forecasting methods include:

• Simple exponential smoothing
• Holt’s linear method
• Winters’ multiplicative season
• Croston’s intermittent demand
• ARIMA (Box-Jenkins)

Many companies use regression models to determine the relationship between demand and demand drivers. They are especially useful for seeing trends and seasonality.

Regression analysis methods include:

• Linear regression
• Hyberbolic trend
• Logarithmic trend
• Square root trend
• Power trend
• Exponential trend
• Seasonal trend

Hybrid forecasting methods combine regression, data smoothing, and other techniques to produce forecasts that can compensate for the weaknesses of individual methods. For example, some forecasting methods are great at short-term forecasting, but cannot capture seasonality.

Hybrid forecasting methods include:

• Vanguard Dampened Trend – a powerful hybrid model that simultaneously detects all trends, cycles, and seasonality in historical data and responds with the most accurate exponential smoothing method. Vanguard Dampened Trend is available across all Vanguard business forecasting applications.
• Log Theta
• Theta

Decomposition forecasting methods are highly effective at finding multiple, subtle patterns. They perform very well at daily forecasts. These methods, additive and multiplicative, decompose historical data into various patterns that can simultaneously include annual seasonality, day-of-the-week patterns, fixed-date and floating-holiday impacts, and others.

Spectral-analysis forecasting models are effective at filtering noise out of cyclical data. For example, spectral analysis can pull seasonal patterns out from data that also have a strong monthly pattern. The seasonality would be otherwise difficult to see.

Spectral analysis forecasting methods include:

• Lowpass filter
• Spectral noise filter
• Formant frequency

When standard forecasting models are unable to capture the way a particular system works, companies will often create their own models to reflect their unique demand signals. These custom methods often are a combination or adjustment to more traditional methods.