In the F&B industry, we put emphasis on procurement and purchasing, however the problem is the traditional method we use to order.
When the procurement is inaccurate in the quantity purchased it could cause: -The restaurants are not be able to produce the requested dishes to the customers due to the lack of ingredients, which in turn would decrease their revenue. -Due to the misjudged order quantity submitted by the restaurant, suppliers would have to replenish the required items raised by the customers which would increase the delivery cost.
That's why we created Ebuy-cast, it offers a revolutionary method of smart ordering system by introducting MLP and Neural network forecast paired with a beautiful easy-to-use user interface.
What it does
Having previously worked in the F&B industry, we know that the day of the week matters when it comes to customer volume. This has a direct impact to the sales data that each restaurant actually gets. By using time series prediction models, we are able to factor in not just a particular day of the week, but also the festive days in the year that will result in anomaly surge in sales. The model is built concurrently with strong data analysis to support our decisions at every step, it recognizes the pattern fluctuation and we are able to observe clear distinct patterns formulated by week/month/year. We also used the prophet model to assist us to detect the Change Points in time series data and regularise the parameters by means of Bayesian oprimisation with cross-validation. We are also able to incorporate the multiplicative-seasonality and determine the uncertainty intervals in the data, thereby producing a meaningful and accurate prediction.
How we built it
Using online open source code reference, we decided to use the best fit ML prediction model that fits our use case, by using EDA + Prophet + MLP neural network forecasting
Challenges we ran into
The ML and Dashboard was developed within 24 hours by four developers. The prediction model posed many challenges such as predicting the orders within 5% error. The forecasting of December 2020 for 12+ items at 3 different restaurants usiing 3 years of data proved to be statistical challenge and finding which model to use to get the most accurate predictions. As of now, we found that Prophet and Neural Network MLP proved to be a good model for prediction.
Accomplishments that we're proud of
We accomplish this hackathon as a team throughout the night and manage to successfully compile a working ML model, coupled with web devolpment.
What we learned
We learn to use various ML model through open source code and implement it according to our use case. We manage to explore r coding and python notebooks and also ultimately decided on the best algorithm to create our prediction model.
What's next for Team Innovator Ebuy-cast
We plan to further develop our web dev which was halfway done, and also to improve the accuracy of our model by using more variations of prediction / regression ML models