A multi-model methodology for forecasting sales and returns of liquefied petroleum gas cylinders

In the liquefied petroleum gas (LPG) cylinder business, one of the most important assets is the LPG cylinder. This work addresses the asset acquisition planning for the LPG cylinder business of a company from the energy sector which has recently started this activity. In order to make the acquisition plan, it was necessary to forecast the sales and the LPG cylinder return rate. For that purpose, an ensemble method using time series techniques, multiple linear regression models and artificial neural networks was employed. Sales forecast was obtained using time series techniques, in particular, moving averages and exponential smoothing. Then, forecast of bottled propane gas sales and return rate was also addressed through multiple linear regression and artificial neural networks. A probability density function was defined for each of the different approaches. Afterward, using Monte Carlo simulation, the forecast values are obtained by a linear combination of the probability density functions, thus producing the final forecast. Results show that the company’s expectation of growth is larger than that predicted by the proposed methodology, which means the company should reflect on its current asset acquisition strategy. By combining different approaches, the proposed multi-model methodology allowed to obtain an accurate forecasting, without requiring a lot of historical data.