Time Series Forecasting uses historical data to forecast future performance, there are a number of methods each with a unique algorithm producing a forecast from a ‘simple linear plot’ attempting to bisect historical results and forecast the trend into the future based on a ‘line’, to more complex algorithms capable of identifying seasonality trends, outliers (there are several different types) and automatically adjusting outliers to create a more accurate forecast and the most involved methods that allow the introduction of external variables, for example weather or sales promotions to further ‘tune’ the forecast results.
The use of basic Time Series Forecasting methods, for example Simple Linear can be invaluable to an organization. Which Time Series methods that will prove consistently accurate in any given organization will vary based on that organizations business objectives, the types of trends in their historical data and most importantly the consistent accuracy and cleanliness of that data.
Business Objective of Time Series Forecasting:
The most important step in Time Series forecasting is determining what is the business objective for using Time Series? The answer will inevitably dictate the approach, Time Series methods and what data and level of data should be forecasted. For example;
- Are you trying to reduce inventory?
- Are you trying to forecast volumes?
- Are your trying to forecast daily data? Monthly data?
These are a few of the possible business objectives, but each has its own implications. If you are trying to forecast sales, is sell in / sell out important? If so, it is likely each has its own trends and how does each impact inventory and which inventory? Manufacturer or the distributor inventory? The emphasis here is the business objectives and measure(s) of success need to be thoroughly identified, and this will dictate how and on what data Time Series forecasting approaches and methods should be applied.
How much data?
The typical recommendation is to use a minimum of 36 rolling months of data (to capture seasonality trends), but in many cases you will find the data does not have 36 months of history (product may not have been around for 36 months), the data contains null data (missing data points) or there is significant variability in the historical data. This is where defining business objectives is critical to determining the approach. Using a Time Series application that can differentiate null and zero values, has advanced capabilities such as automatic outlier detection and correction, supports Time Series methods optimized to generate forecasts based on limited historical data (new products) are absolutely required for success.
Data accuracy and cleanliness
The intent of Time Series forecasting is to detect trends in historical data over time and forecast future periods. If the data is not accurate, inconsistently accurate or requires prior period adjustments this will affect the accuracy of the forecasts. The business objectives and measures of success will dictate which data will be required, once this is determined a thorough ‘honest’ assessment of the accuracy and cleanliness of historical data needs to be completed. Otherwise you risk creating forecasts that are simply inaccurate which could lead to poor business decisions.
Measuring Time Series Forecast Accuracy
Two common metrics to assess the accuracy and effectiveness of Time Series forecasts over time is Bias and Forecast Accuracy. Bias is the percentage by which your forecasts are incorrect, Forecast Accuracy is the absolute forecast percent accuracy. Both metrics should be assessed periodically and over time, based on the business objectives and measures of success. For example, if your objective is to accurately forecast sales 2 months out to manage inventory your Bias and Forecast Accuracy measures should be, at a minimum on a 2 month rolling cycle.
The Time Series Forecasting Cycle
To effectively forecast using Time Series methods requires a commitment to the entire Time Series Forecast Cycle;
Time Series Forecast > Measure Accuracy > Adjust Forecast Algorithms > Time Series Forecast
The results of each forecast should be assessed by period (ie month) and over time, if the forecasts are not accurate or producing an unacceptable bias – why? Should a different approach for that specific forecast (ie Product) be considered? Different method? Are there data issues?
A benefit of Time Series forecasting is the ability to reforecast history to determine if a different approach or method would have produced better or worse results. Time Series forecasting should not only be applied to the current period to produce a future forecast, but should be used as part of the Time Series Forecasting Cycle to test and validate assumptions on alternate approaches and methods with the objective of always trying to improve the forecast accuracy.
Selecting an application – Minimum Functional Requirements:
Before selecting an application, we recommend determining the business objectives and measures of success which will dictate;
- Time Series forecasting methods required
- Is null data relevant (almost always)
- Types of outlier detection and outlier correction (automatic)
- Introduction of Independent variables to improve forecast accuracy
- Amount of data
Time Series models are highly complex algorithms, “math problems” that require an application to scale in both processing power and ability to handle large volumes of data, confirm the application can support your organizations volumes.
Below are functional requirements considerations:
- Null vs Zero values – in Time Series Forecasting, a null value is typically, and should be treated differently than a zero value. A zero value is just that, the data point is zero as in 0 Revenue, 0 Expenses or 0 Products Sold. A null value is a data point that doesn’t exist, for example a missing data point or a period where a product wasn’t being sold which should be interpreted differently by the Time Series Algorithm than a 0 value. Treatment of a null value as a zero will inaccurately skew the Time Series Forecast.
- Automatic Selection of best forecast method – the more advanced and functional statistical applications will execute multiple Time Series Methods on the same historical data and select the ‘best fit’. Understand how the application automatically selects the best method. Is MAPE (Mean Average Percentage Error), the average error over time the only approach for selecting the best method? If the MAPE is 10%, they your forecast could be expected to be wrong by 10%, is this acceptable? What about other ‘goodness of fit measures’ for selecting the best method?
- Outlier detection and correction – it is important to understand there are several different types of outliers, for example a local or single period outlier vs a fundamental shift in the historical data (up or down). Can the application both detect and automatically correct different types of outliers? Does the application support user defined identification of outliers and user defined rules to automatically correct? The risk of not supporting this functionality is a higher probability the forecasts will be incorrect and/or extensive manual effort to adjust outliers.
- Independent Variables – Some Time Series methods support the introduction of independent variables to further ‘tune’ a forecast, for example weather, sales promotions etc. Does the application support these methods? Support the ability to easily include independent variables?
As noted above, the most basic use of Time Series forecasting can be invaluable to an organization, embracing the Time Series Forecasting Cycle can increase revenues, reduce expenses and optimize processes including the forecasting cycle itself, inventory, manufacturing and sales operations.
Check back for future posts on Time Series Methods, Measuring accuracy and Best Practices.
….this is not an apple