Skip to content Skip to footer

The great AI frontier: time series

Image: Adobe Stock / khunkorn

What do the following four questions have in common?

  1. Will Wrexham or Notts County win the National League this season?
  2. Will Tom Brady unretire?
  3. Will my friend conquer Stage 3 bowel cancer?
  4. What will Man Investments share price be next Thursday?

An answer: ChatGPT cannot give me a specific answer. ChatGPT responds per, “as an AI language model, I do not have the ability to predict the future with certainty, and I do not have access to information beyond my training data.”

In the future, will AI be able to answer these questions? Well, for two of the above questions, ChatGPT responses were already useful. It provided helpful cancer guidance for my friend, plus a balanced response on Notts versus Wrexham.

While not giving a direct answer, the AI – represented by ChatGPT – did great! It understood the questions, gave a well-written and reasonably substantial responses for two. And for the last question it recommended getting advice and conducting intense research on Man Investments stock price. Can and does AI do more beyond ChatGPT? Yes, absolutely. And we’ll return to Tom Brady’s retirement too!

Why can AI beat chess grandmasters and go champions, but not offer simple predictions to my questions?

AI, and more specifically the machine and deep learning algorithms underpinning much of it, is already proven in so many ways. Highlights included IBM’s Deep Blue beating Garry Kasparov at chess way back in 1996, while 20 years later, DeepMind’s AlphaGo began to beat champions in the popular game of Go!

But while impressive feats both mathematically and philosophically, neither come close to the difficulty of predicting Wrexham or Notts County’s success, Man Investments share price, my friend’s likelihood of conquering cancer, or Tom Brady unretiring (again). Why?

Well chess and Go have finite possibilities, a maximum number of potential sequential moves. That means a great deep learning model can be trained – with sufficient compute – to predict optimal moves all the time over a finite number of board positions. If a brilliant human wins, the model learns, so it won’t lose again. Making it a very big data problem, but a finite one. Hence victory for Deep Blue and AlphaGo came with the right models and sufficiently powerful compute, the latter accounting for the 20-year gap between them as Go has significantly more possibilities than chess.

However, the battle between Notts versus Wrexham, like the stock market and the health recovery of my friend, has infinite ‘moves’ and possibilities. As ChatGPT states: “Both Wrexham and Notts County will need to perform consistently well throughout the season to have a chance of winning the title. Ultimately, it will come down to factors such as team form, injuries, and luck.”

Accuracy metrics can also help manage predictive uncertainty. For example, my water utility knocked on my door recently to tell me their bot had predicted a sizeable leak in front of my house. The company digs up the road, and the bot was right! But the bot could have also been wrong.

False positives are a frequent dilemma of machine learning, potentially wasting time and money in digging up the road. For my water utility, 99% accuracy is better – and less costly in terms of incorrect detections and wasteful digs – than 90% accuracy, which in turn beats 50%. Humans can step in to validate or check the false positives as required. AI delivers real results, even with its imperfections, through insight if not perfect prediction.

And time series?

All the problems outlined thus far can be set up as time-series, patterns over time represented as vectors, columns and matrices. The sequencing of moves in chess, the utilities leaks datasets, the medical databases about treatments and responses, the Man Investments share price (and accompanying market factors), and seasonal fortunes of Wrexham and Notts County, all time-series.

This is where time-series analytics comes into its own, and the ability to join the most granular data-sets, and to analyse them with relevant model sets. Regression methods of multiple types can be used to model the data in order to make predictions. With particular focus on time series oriented neural networks and deep learning methods such as long-short term memory (LSTM) neural nets, reinforcement learning, recurrent neural networks. All are incredibly powerful model types.

Take, for example, the recent news article on Morocco’s World Cup run – and the man who helped plot it. Harrison Kingston, a data scientist, who became Director of Performance at the Moroccan Football Federation refers to not being able to unlock a ready-made route to goal and success no matter how much data and performance analytics, powered by AI, is utilised. However, he can maximise the possibility.

In essence, he states a time-series problem and opportunity. He makes as much sense of infinite possibilities in time for a win every time. And in this case, what is true on the pitch applies off of it, in healthcare, in the stock market, and predicting water leaks.

For real-world decision-making, AI and its most current exciting incarnation – ChatGPT – can provide direction and context. But that in itself is not enough – insight and context needs human interpretation. And that interpretation comes from the environment that allows fast querying, versatile modelling, automated anomaly detection, and insight visualisation, best enabled by a powerful time series capability.

As one industry leader put it about his time series analytics engine, “we see no limits to all conceivable data in the universe, capturing it in real-time to record a ‘continuous stream of truth’.” ChatGPT-like, “we ask any question of any dataset at any time to get an answer instantaneously.” Taking the ChatGPT analogy further, “the only limit is our imagination and in the questions we can think of to ask … such as why did something happen? Or why did something happen the way it happened? Or what were the small events that led to the big event? And which factors influenced each event? If you know the answer to the question ‘why’…”, the world is your oyster.

And thus, his organisation gets AI observability nirvana.

What about Tom Brady unretiring?

ChatGPT responded, “I don’t have access to Tom Brady’s thoughts or future plans. However, as of my knowledge cutoff date of September 2021, Tom Brady was still an active player in the NFL, playing for the Tampa Bay Buccaneers.”

A lot happened since September 2021, and who knows what could yet happen?

Picture of Steve Wilcockson
Steve Wilcockson
Data Science Lead at KX

You may also like

Stay In The Know

Get the Data Centre Review Newsletter direct to your inbox.