RSM South Africa

Machine learning and implications for accounting

“The future is here, it’s just not evenly distributed.” – William Ford Gibson.

It is fair to say computers have had a profound impact on the way accountants work. The last three or so decades have seen the spread of accounting software, automated controls, paperless processing systems, cloud accounting, and computer aided auditing systems. Could the industry be about to see a much bigger revolution with the advent of machine learning?

In the article that follows, we briefly explore what machine learning is and why it might have a profound impact on the accounting world.

What is machine learning?

To establish why machine learning is so important, it is important to distinguish it from traditional software programming. Most computer processing is done based on specific instructions that someone wrote. The programmers have to clearly define what inputs the system can accept, and what it should do with them. This works well in a wide range of scenarios, but it has inherent limitations. This is because the systems can only do what they were explicitly instructed to do.

Machine learning, on the other hand, is an attempt to replicate the way human brains work.  Programming in this model consists of training computers by providing them with input data and desired outcomes, and letting them develop statistical inference and pattern recognition to develop connections that would be necessary to go from input data to desired output. 

For a basic illustration of how this works, we can use an example from the HBO show, Silicon Valley. Without getting into the plot details, some characters developed an app called Not Hotdog that had one job: to identify whether the image it was looking at was a hot dog or not. Now, while this is a silly use case without much practical application, it would be a very difficult computer programming challenge using traditional programming methods. Using machine learning, you feed the system with thousands of images, with some labelled as ‘hotdog’ and others ‘not hotdog.’ The system develops its own pattern recognition systems to determine what makes something a hotdog. 

As a side note, the app that was used on the TV show was a fully functional app. As reported by Techcrunch, the engineer who made it for the show used a dataset of 150,000 images. A majority, 147,000 images, were not hotdogs, while 3,000 of the images were of hotdogs. This ratio was intentional to reflect the fact that most objects in the world are not hotdogs.

Recent advancements in neural network processors, which are computer chips designed to work in a similar way to the human brain, have seen strides in machine learning applications. Google’s AlphaGo system has beaten the champions at the highly complex Chinese game of Go using strategies that no human had thought of. The system basically taught itself to play the game.

Beyond hotdogs and Go, this technology is being used for tasks as mundane as making the pictures your smart phone takes much better, to complex challenges like natural language processing and self-driving cars.

So what does this have to do with accounting?

Current accounting software makes processing of transactions much faster, easier and more accurate than manual processing. However, they have some hard limits. They cannot, for example, exercise judgement and make decisions. This means tasks which require any decision making have to be performed by humans.

With machine learning, software can be trained to make decisions. This could automate many accounting functions, from calculating and processing journal entries to preparing financial reports. 

While it may seem unlikely that computers can exercise judgement, most judgement in accounting is based on probability of different outcomes. This is a key strength of machine learning systems as long as they are trained on sufficient data.

A recent development is adversarial networks, where one system produces results and another evaluates them. Both system are then given feedback which they use to fine-tune their algorithms before starting the process again. There is potential for this to be used to train accounting systems, where one system processes accounting transactions and another audits the results.

The result of this could be much faster and more accurate transaction processing, period end processing, financial statement preparation, analysis and assurance. This should all be possible once existing technology has been adapted to the accounting industry. 

This is not to say humans will be taken out of the system. Financial reporting is still an arcane art to most people and just giving them the output produced by some advanced computer system is probably not going to satisfy their decision making needs. In this scenario, accountants will primarily be in the advisory business.

This raises some interesting questions about liability in the case of inaccurate results, which may hold back the adoption of this technology in the finance industry. This is because no one actually writes the code in a machine learning algorithm, the computer writes its own code based on the training data and feedback it has been provided. It will also be interesting to see how accounting and auditing standards will evolve in response. As with any new technology, there will be teething pains.

Leonard Mundida

Supervisor: Audit, Johannesburg


Related articles

IT meets audit   

RSM's top cybersecurity tips