Kurvv’s AutoForecast product provides customers with a variety of quantitative sales forecasting methods, so they can simply connect their data source and receive a customized, accurate forecast in seconds. Once data is uploaded, AutoForecast tests out several different time-series forecasting methods including decomposition, exponential smoothing, ARIMA and regression (see below for more details about each). Finally, an average of all forecasts is computed and output with the final results. Customers can select the forecast output(s) they feel most comfortable with and customize forecast time horizon and validation time windows.
Any quantitative forecasting method relies on two key assumptions:
If either of those assumptions is not true for your business, then quantitative forecasting methods such as Kurvv’s AutoForecast product should not be used. However, Kurvv can help you custom develop a forecast using qualitative approaches – just reach out[here] to learn more!
*as of the date of writing (4/14/2020) the world is going through the COVID-19 pandemic and nothing, including sales, is “normal”. Given these abnormal times, any models trained with past historical data from “normal” times will probably not be able to predict sales with any accuracy during these times, and models that will be trained with current “abnormal” data will probably not be able to predict sales with any accuracy as well.
It’s also worth noting that most generic forecasting products can be improved with customization for a particular customer, either by supplementing models with additional metadata specific to that business or by tuning existing model features and parameters based on individual performance characteristics.
Finally, Kurvv’s AutoForecast product is designed to predict total bookings or total unit sales by day off the shelf – if you need more granular predictions – say, sales by room type or product category, or if you’re interested in a more customized approach for any reason, contact us[here]!
As noted above, AutoForecast starts with data customers either manually upload or connect automatically through our supported integrations[here]. From there, we overlay proprietary metadata about seasonal events and test out the following time-series* forecasting methods.
Linear regression is a statistical method often used for prediction which seeks to use external information (independent variables) to describe the behavior of a dependent variable such as revenue or units. When using regression to forecast some dependent variable over time, regression is also considered to be a time-series forecast method. Kurvv’s regression forecast overlays information about holidays, days of week and a variety of other seasonal descriptors to see how well seasonality describes a customer’s forecasted metric.
Time-series decomposition is almost self-defining – the models attempt to extract from your data information about long-term trend, seasonality and cyclical behavior. If the model is successful, these components can be extrapolated to future dates to produce a forecast.
Exponential smoothing is one of the most widely-used time-series methods, and basically works by averaging trailing data while weighting more recent results more highly than less-recent results when projecting future dates. Because most ecommerce and hospitality businesses have seasonality, Kurvv uses the Holt-Winters’ seasonal method (Reference: Rob J Hyndman) which captures seasonality.
ARIMA stands for auto-regressive integrated moving averages. It is another commonly used forecasting approach which is similar to regression – but instead of using observations of other variables to predict an outcome, autoregression uses prior observations of the predicted variable and lagged effects to predict future values.
AutoForecast creates a daily forecast for your data using each of the methods above plus a fifth method of averaging each forecast to provide the final output for the next 6 months or your specified future horizon. In addition, summary statistics and predictive performance are provided for each forecast model and for a user-defined lookback validation window.s