Predicting the outcome of a general election is a challenge. But combining quantum computing with neural network technology could improve forecasts, according to a new study that used just such a network to model the 2016 US presidential elections.
鈥淯sing this application is a way of demonstrating the potential of these new sorts of algorithms, and the kind of problems that quantum computing might be able to assist with in the future,鈥� says Michael Brett, CEO of data analytics company , which conducted the study. Quantum computers use 鈥渜ubits鈥� instead of standard bits, which can be in a mixture of two states at once, vastly increasing their processing speed.
Most polling organisations incorrectly predicted that Hillary Clinton would win the 2016 presidential election. Many analysts concluded that one reason for the inaccuracy was that the predictive models used did not take into account the fact that individual states are not isolated and correlations between them can have an effect.
Advertisement
That鈥檚 because 鈥渄eep鈥� neural networks capable of factoring in such correlations are difficult to build. The budding field of quantum machine learning could one day provide a faster, easier alternative as the devices reach their full processing potential, according to Maxwell Henderson, who led the QX Branch project.
Weighting states
His team chose a type of deep learning network made up of many interconnected nodes; the more nodes, the deeper the network.
Modelling the 2016 election is a simpler task compared with most big data analytical projects, with just 51 nodes, corresponding to the 50 states and Washington DC. Each node can be one of two possible configurations, corresponding to a Democrat or Republican outcome for that state.
Each state is connected to every other so they can interact freely and signals can travel between them. The states also each have their own bias towards one outcome. The strength of the interactions between the nodes depends on the voting correlations between states in past elections 鈥� those that tend to vote together will interact more strongly in the model. The sum of all those biases and weights combine to influence the probability of a given outcome.
The QX Branch team used the controversial D-Wave quantum computer to train the network 鈥� drawing on polling data collected by the forecasting site Five Thirty Eight and results from the past 11 presidential elections. They then ran simulations of the 2016 presidential race.
The network confirmed that crucial swing states were strongly correlated to the election outcome 鈥� that is, they were more likely to influence the final outcome, compared with reliably deep red or deep blue states like Alabama or California.
Overall, the quantum neural network arguably did slightly better than Five Thirty Eight in forecasting the 2016 election outcome. It predicted a roughly 50:50 toss-up race, with a slight Trump advantage 鈥� close enough to reality聽that it qualifies as a strong result.
Taking the advantage
But the work is more a proof of principle. The study only included 49 of the 50 states, omitting Maryland and Washington DC due to limited computing power. If the D-Wave computer improves enough by 2020 the team hopes to test the model again in that election.
Higher powered machines are not yet available because quantum computing is still very much in its infancy. And according to at the California Institute of Technology in Pasadena, it鈥檚 unclear that using quantum machine learning offers a significant advantage over classical computing in this instance.
鈥淭his is a very small problem with only 50 states, so you could use a classical computer to replicate everything in this study in a pretty rapid time frame,鈥� says Crosson. But she thinks this kind of deep learning network run on a regular, non-quantum computer could still useful for more accurate election forecasting.
Henderson remains sanguine about the promise of quantum computers. 鈥淚t鈥檚 not going to be a stand-alone, solve everything application,鈥� he says. 鈥淨uantum computing is not going to overtake Five Thirty Eight鈥檚 modelling work overnight. But it could help with the difficult problem of modelling correlated systems, and maybe steer forecasters to also experiment with some of these models in the future.鈥�
Read more: How YouGov鈥檚 experimental poll correctly called the UK election
Reference: arXiv,
