Harvested holistic demand intelligence:
Demand is shaped by both market sentiment and business-driven variability caused by marketing promotions. Forecasting methods can gauge demand with a higher degree of accuracy when all variability factors are integrated. We extracted historical sales data to analyze causal relationships between sales and variables - including season, week of the month, day of the week, time of day, holidays, festivities and end of season promotions. We analyzed demand at the master SKU & aggregated periodic levels to identify the most suitable statistical method and time horizon for forecasting demand trends with minimal forecast error. Exploratory Data Analysis (EDA) presented a great way to break down datasets and construct this predictive model.
Transformed insights into forecasts:
We adopted multiple time series models to ensure all relevant parameters, including complex permutations such as channel and category, were considered at every level of forecasting. We also created plots in R and Tableau to distill insights from EDA, and enhanced the accuracy of modeling. We used Auto.Arima, STLF and ETS functions in R to generate point forecasts, actual forecasts, and the mean of predicted values, and the TBATS time series decomposition method to measure diverse components and decipher patterns. We selected the most appropriate time series technique to forecast multi-dimensional data, and included Ensemble machine learning algorithms to ingest new data points and improve the credibility of forecasting.
Built informed business plans:
We built a plug-and-play demand modeling tool to deliver business intelligence for operations planning. The dashboard of current sales and future demand enabled merchandise planners to view channel-wise traffic, along with the spikes and troughs in demand, anytime, anywhere.