Twitter has announced it will introduce updates to prevent tweets from disappearing when a user’s timeline auto-refreshes. In a tweet posted on 22 September,…
The challenge for restaurant managers
Founded in 1997, Hungry Lion is a market leader in the fried chicken fast food sector with 230 stores across Africa.
“We simply didn’t have enough business intelligence to predict sales on a particular day for a particular store,” says Gideon Jacobs, Business Development Manager at Hungry Lion. “We have 230 stores across multiple countries, all with varying sales figures. No person can predict that accurately – but a machine algorithm can. It looks at the historic values, actual performance, as well as behaviour and economic insights to predict sales volumes.”
Hungry Lion ran a ‘man vs. machine’ competition and compared the results
Hungry Lion restaurant managers juggle many duties, such as determining customer demand, ordering stock, reducing waste, checking on customers, managing and scheduling staff. Staff scheduling takes significant time and requires accuracy to ensure customer satisfaction and continuing profit through the pandemic.
Projecting sales volumes and scheduling staff is usually done through a combination of instinct, experience and pure guesswork by managers. Predictions were not based on solid data; it was a rule of thumb, resulting in overall demand forecasting errors of 24%, based on a specific study done with managers during a controlled test run.
Branch managers were off by more than 40% for every one in five sales predictions made. Some managers have a better success rate than others. Nonetheless, a large part of their job is to anticipate how busy they might be on any given day and then schedule the right number of staff members for a successful day of trading.
If there are too many people on duty, staff stand idle. If there are too few, they are unable to cope with demand, resulting in overworked employees and unhappy customers. This also makes for inaccurate stock ordering; stock levels are typically matched to predicted sales volumes so a surplus results in wastage or a shortage, again, leads to unsatisfied customers.
Combining human and machine intelligence to improve accuracy
Hungry Lion brought in Predictive Insights to compliment the human element with machine learning algorithms that reduce prediction inaccuracies. The startup developed a process that extracts point of sales data, staff scheduling data and clock-in data (showing the time staff arrived and left and when they took breaks) from each of Hungry Lion’s branches daily. The system then checks the data for any anomalies including peculiarly slow sales, missing or incomplete clock-in data and incomplete schedules.
The system communicates this to Hungry Lion and imputes reasonable values for the data where necessary. Finally, it combines this with historical data and a curated dataset containing information such as SASSA payment dates, instances of load shedding and weather.
The algorithms combine machine learning methods and economic models of consumer behaviour, calibrated to aggressively learn about any new trends in behaviour, including changes expected in various COVID-19 lockdown levels. The data is used to predict daily sales for each branch for the next five weeks.
Demand forecasting errors cut by 40%
As with any change process, Hungry Lion was hesitant. The chain ran a ‘man vs. machine’ competition and compared the results. The machine did overwhelmingly better.
Branch managers were asked to predict sales two weeks in advance. This was compared to Predictive Insights predictions over several weeks. Out of 92 branch managers who submitted forecasts, the average forecast error was 24%. Predictive Insights’ machine forecast error margin was 15%.
The machine was particularly effective at avoiding large forecasting errors. Branch manager forecasts were out by more than 50% for 13% of the time, whereas this only occurred for 2% of the machine’s forecasts.
Cost of scheduling errors reduced from 34% to 20%
Sales predictions and labour scheduling rules are combined to create recommendations for the number of workers scheduled for each day. Long-term forecasts are used for internal financial planning and to set sales targets for branches.
The cost of scheduling errors, when expressed as a share of total wage costs, was reduced from 34% to 20%. This implies a reduction of wasteful spending equal to 14% of the wage bill.
Predictive Insights produces reports on historical labour scheduling errors to quantify the financial costs of worker absenteeism and scheduling mistakes by branch managers. This information is fed back to regional and branch managers, creating accountability.
“Artificial intelligence reduces the risk of branch managers trying to accurately predict sales, stock and staff volumes, especially during turbulent times,” says Gideon. “Our margins of error have improved for demand forecasting by 40%. Now we can get pretty close to the right number of personnel hours needed because the predictions are much more accurate.”
Hungry Lion has already reported a decrease in periods of being either under-or overstaffed, aligning the wage bill more accurately to demand and benefitting their bottom line, staff, managers and most importantly, their customers.
Featured image: Ke Vin via Unsplash