We recently spoke to Helgi Helgason, our Chief Technology Officer, about his recent work on our upcoming sales prediction model. We’re really excited about the potential this might hold for our clients and the future of the Live Entertainment in general. It really could be a game-changer, combining the most cutting edge data analytics with our unique position in the industry and access to data.
Our CTO, Helgi Helgason, somewhere off the coast of Spain. While not working on the prediction model.
Here’s a quick look behind the scenes at what we’re working on. Hopefully, Helgi can clear up some of the confusion (or as much as you can expect from a tech expert). For anyone struggling, don’t worry, just read our ‘human-level’ summaries!
AI-based predictions. Sounds very advanced. Can you tell us what it is?
Helgi: “Basically, the idea is that we can draw upon all of the historical data that we’ve accumulated in order to predict sales for future events. We apply deep learning to historical data, where a neural network learns complex patterns and correlations that are then used to make predictions.
Since Activity Stream is partnered with different organizations in the live entertainment industry from around the world, we have access to a very large (combined) dataset which gives the potential for a very powerful model.
With deep learning, prediction models can understand non-linear patterns which would not be possible using regular statistical techniques. One direct result of this is that our sales predictions capture nuances such as slower or faster sales during a specific part of the sales period. That leads to insights such as when might be a good time to increase or decrease marketing spend. Our model works with over 150 input variables and produces a predicted sales curve by leveraging the learned relationships of the inputs. In the end, our customers see accurate estimates for total sales and how sales are likely to unfold over each part of the sales period.”
Human-Level Summary: Big computer, many data points, won’t fit in Excel. Or your head.
How is it different from looking at a sales graph and extending it to make a prediction?
Helgi: “A simple prediction from a classical sales graph will typically only be based on past sales and is created using linear methods such as regression. Using linear techniques means that any predicted sales curve will essentially be a straight line, which is unlikely to be an accurate prediction. Things tend not to be linear in the real world…
Visualization of Neural Networks
This might seem a little complex, but the most important practical difference between linear regression and deep learning models is ultimately the capacity to deal with “curvy” patterns and how they perform when the amount of data increases. Where there is a minimal amount of data, deep learning approaches tend to actually perform worse than traditional approaches. However, when faced with a large data set, they are able to associate the data in a much more nuanced way.
This is something that is massively important in a project like our sales prediction model, which draws upon over 150 different dimensions in formulating its predictions.”
For Humans: More power means more nuance, hence higher accuracy.
What data are you drawing upon in order to create these sales predictions?
Helgi: “Specifically, the types of data that our sales prediction model will consider are things like how customers are interacting with the product (purchases and more), customer demographics, product attributes, and location-related factors.
In general, Activity Stream is able to gather all of this because of our unique access to a massive amount of data. This means that we are able to approach sales predictions in a wider sense, not just being reliant on data from one venue or organization. Coupled with our ability to work with deep learning, this creates a perfect situation for us to deliver quality predictions to our customers.”
For Humans: Loads, from different clients.
Will it require organizations to be working with Activity Stream for a long time in order to get the model to understand patterns and make qualified predictions?
Helgi: “No. Customers will be able to access the sales predictions from day one, as soon as they have been launched with Activity Stream. Over time the prediction model improves as it is trained on increasing amounts of data, which accumulates over time as customers use our solution.”
For Humans: No.
How accurate will it be?
“Based on measurements from our prototype, the average error of our sales prediction model is less than 5%. Those results are promising and we aim to reduce this further as we train the model with more data and make other tweaks. Testing is carried out by taking a part of the historical data out of the training data set so we can measure the performance of the model on data it has never seen before (which is the normal use case).”
For Humans: Over 95% currently.
In Conclusion…
There’s a lot of excitement here about what our new sales prediction model is going to provide for our customers. Up until now, there haven’t been any sales prediction models in the ticketing industry that deal with this depth or breadth of data.
Hopefully, this article made our new project a little simpler to understand, or as much as we could hope when trying to explain machine learning models.
We’re looking forward to seeing what our existing clients can get out of this new model as soon as we pilot it in early 2020. We’ll let you know the results as soon as we do!
For Humans: We’ll let you know when we launch. We promise.