In the dynamic world of luxury hospitality, effective demand forecasting is crucial for a leading hospitality chain that oversees a diverse portfolio encompassing hotels, resorts, jungle safaris, palaces, spas, and in-flight catering services. The company embarked on a demand forecasting initiative with the following objectives:

  • To generate item-level (SKU) forecasts for each unit, i.e., property or hotel within their extensive network.
  • To provide monthly-level forecasts for a 3-month period, facilitating informed procurement and inventory management.
  • To prevent stock-outs and reduce wastage of items by ensuring the right quantities are available at the right time.
  • To minimize packaging and transportation costs through improved demand prediction.


To address these objectives, the company adopted a systematic approach that involved:

Grouping of Forecast Series: Over 10,000 forecast series, a combination of hotel properties and items, were categorized into 12 buckets based on consumption history. The selection of the forecasting strategy/methodology was contingent on the length of consumption history and the frequency of consumption.

Forecasting Techniques: Various forecasting techniques were employed. For series with a substantial 48-month data history, models such as ARIMA, Holt-Winter’s seasonal technique, and CatBoost-CARMA were employed. In the immediate short-term, a 3-month weighted moving average (3WMA) was recommended, particularly in light of the disruptions caused by the Covid-19 pandemic.

Model Building and Performance Evaluation:

  • We used 45 months’ data to train the model (JulyXX to MarchXX)
  • We then forecast the consumption for next 3 months (AprXX – JunXX)
  • Model performance evaluated by comparing forecasts to actual consumption using MAPE (mean absolute percentage error) as the performance metric
  • The model performance is then compared to baseline forecasts (Weighted Average, Naïve & Seasonal Naïve)


  • Automation: The development of automated scripts capable of ingesting and processing data efficiently, generating accurate forecasts for a 3-month horizon.
  • Optimal Model Selection: The identification that ARIMA-based forecasting models excelled when 48 months of historical data were available. For datasets with less than 48 months of data, the performance of ARIMA was found to be comparable to 3WMA.
  • Cost Savings: A comparison of procurement costs during the period of April to June 20XX revealed that the ARIMA model outperformed Seasonal Naïve (+15% Trend) for forecast series groups with 48 months of full data, accounting for 33% of the total procurement value in 20XX HX.
Want to learn more about TransOrg’s value proposition, solution methodology and implementation approach?