Trend

Trend, combined with seasonality and demand lags, forms a core structural characteristic of a time series. Typically, the trend is assumed to be linear in nature, with the slope co-efficient remaining constant across all time periods. While this is true for some some time series, the presence of cyclical, non-linear, or localized trend with varying slope over time should not be ignored when projecting demand into the future. Also, the presence of level shifts can adversely affect the quality of the model used for identifying trend, and this in turn this can interfere with the fitting procedure of several machine learning algorithms.

To overcome limitations associated with strictly assuming a parametric (or analytical) formula for estimating trend, Demand Guru utilizes a hybrid parametric/non-parametric model that can adapt to local changes in trend, while remaining sensitive to the presence of level shifts and other outliers present within the data. This hybrid approach automatically distinguishes between:

• a parametric model, when an appropriate closed form model can be fit with acceptable quality
• a non-parametric localized model, when a closed form model yields an inferior fit indicating that the data is affected more by local trends rather than a global one.

The trend model is validated at every step to ensure an acceptable quality of the extrapolation. The resulting trend, when extracted, is input to the application's machine learning algorithms as a causal variable.

However, one of the key drawbacks with extrapolating trend purely based on historical data is that it can ignore causality. For example, what if a product is experiencing growth in demand, such as the sales of a luxury item due to an improving economy fueled by lower interest rates and lower tax rates? Projecting that the demand is going to continue to rise, while ignoring the causal factors ( the tax and interest rates and the improving economy) can severely affect the quality of the trend that is projected into the future. To avoid such issues, and to exploit causative modeling to its maximum possible effects, Demand Guru does the following:

• After extracting the trend from the data, it is compared with causal factors that the user has provided.
• If a significant correlation is detected between a causal factor and the extracted trend, the trend extracted by the hybrid model is ignored, and the causal variable is utilized to project the trend into the future.

This method also helps you construct scenarios on key indicators/causal variables that could corroborate future demand.