Mystic Megabytes

"Our project was an investigation into the uses of AI technology, specifically Machine Learning algorithms to forecast time series date at scale. The project was inspired by the manual and repetitive forecast process the business currently uses which is highly disruptive and not scalable. During our project we stepped through numerous forecasting processes from around the business to understand the underlying business need, and the inputs and outputs required. Two interlinking factors identified were the desire for accuracy and automation. We were also able to make a clear distinction between a forecast and a target, the current process merges those two views which makes it difficult to distinguish between our underlying run-rate and our progress towards our aspirational goals. We tested and compared different forecasting approaches, namely Facebook's ""Prophet"", ""Neural Prophet"", and ""AWS Forecast"". After comparing the results, we found that AWS Forecast was slow and expensive, while both Prophet and Neural Prophet approaches were fast. However, Neural Prophet was more accurate, making it the recommended solution for our hackathon project as it is accurate and scalable if deployed in real-time. Initially, we only forecasted revenue, but later we expanded our approach to other KPI metrics such as predicting bookings. We observed a similar level of accuracy for this metric, and we believe that we can extend this approach to the rest of the KPIs in the business. Overall, we were able to develop a prototype solution that forecasted our revenue by day, week, and month from the beginning of the financial year to period 5. The results were outstanding, achieving a prediction error rate of just 10.09% and a YTD financial forecast variance of (£+404k, +1%) compared to the budgeted figure which was (£+5,583k, +19%) vs the actual £29.2m achieved. This was achieved through limited fine tuning. We believe we could have further improved the accuracy given more time by informing the model of various known periods such as Covid, and key trading periods and when trading tools have been used (Low Deposit).

The project clearly highlights an opportunity to gain both efficiency and improved accuracy through implementing AI technology. Although we are only in the concept stage if implemented correctly the main benefits to the business can be summarised as :

• Improved accuracy (Remove Bias, Human/Process Errors, more frequent)
• Scalability (Standardised process not reliant on Excel)
• Reduced resource requirements (est. £500-750k pa currently spent on budget and forecast process)
• Frees up time to work on strategic value add opportunities
• Less Disruption to business
• Improved employee satisfaction (forecasting process is tedious and repetitive)
• Impartial, un-biased view of "risk" to target
• Ability to easily add "What if" scenarios (What if we ran low deposit in period X?)