Digital Assets Report

Newsletter

Like this article?

Sign up to our free newsletter

Training an AI model is critical to effective integration

Related Topics

Automated document processing has already progressed far beyond many people’s expectations and the capabilities available provide hedge fund managers with a number of significant benefits. However, the human element remains key, and training an artificial intelligence (AI) program is crucial for investors to take full advantage of this technology’s power.

By Angele Paris – Automated document processing has already progressed far beyond many people’s expectations and the capabilities available provide hedge fund managers with a number of significant benefits. However, the human element remains key, and training an artificial intelligence (AI) program is crucial for investors to take full advantage of this technology’s power.

“Document processing, whether automated or not, is about language,” points out John Ashley, General Manager, Financial Services and Technology, NVIDIA, “The processing of language has been one of the greatest advances we’ve seen in the AI space.” 

From speeding trades with AI to natural language processing, the time to implement AI is now. Progress and capabilities in this area are only expected to grow. Statistics from Mordor Intelligence show the global AI in the Fintech market was estimated at USD7.91 billion in 2020. This figure is forecast to reach USD26.67 billion by 2026.  

According to Deloitte, employing intelligent automation technologies can halve the capacity required from finance functions, which effectively means doubling their output. This would also increase “the intelligence and quality of information.” 

This can be seen in practice at Applica. The fintech firm has deployed progressive AI to streamline text-based workflows that deliver better-than-human performance. Its robotic text automation platform uses NVIDIA GPUs for training machine learning models and inference in production. Here, the use of AI eliminates up to 90 per cent of manual errors, document turnover to less than one second, and reduces physical workforce effort by up to 75 per cent. 

In terms of how this is further applied within the investment management space, Boston Consulting Group’s Global Asset Management Report 2021 suggests that recent advances in data quality and analytical capabilities, such as machine learning and natural language processing, have led to impressive achievements gained from digital distribution. 

Necessary training 

A trader or investment analyst is not physically able to read every regulatory filing or financial report, nor can they read every relevant interview or article across the entire industry segment they cover. However, they could train an AI program to do this on their behalf. The same applies to quantitative and algorithmic traders if they want to incorporate signals derived from documents and other text data.  

Algorithmic trading has several advantages such as automation, speed, accuracy and reduced costs, data retrieval and trading speed, but most importantly not letting human emotions impact trade. Accuracy is key since algorithms are made of complex correlations of data, events, and pattern detection that enable them to react to information including news, market price changes, events, and place trade accordingly to make profit. 

But this process is not just about downloading a tool and setting it to work. 

Ashley outlines: “In a research setting, an automated tool meets, or even beats, the level of human accuracy. In the real world, things aren’t so simple; data is not always clean, writing isn’t so clear if the author is trying to hide something, for example. In many cases, these discrepancies and gaps are the parts which are of greatest interest to hedge fund managers.” 

Keeping people in the loop 

“There is a growing research body of Human-in-the-loop (HITL) natural language processing (NLP) frameworks that continuously integrate human feedback to improve the model itself. HITL NLP research is nascent but multifarious [sic] — solving various NLP problems, collecting diverse feedback from different people, and applying different methods to learn from collected feedback,” write Zijie J. Wang, Dongjin Choi, Shenyu Xu and Diyi Yang, from the College of Computing, Georgia Institute of Technology[1]. 

In Ashley’s view, the AI model should be considered a smart intern showing up on their first day of work: “These models learn from data which is fed into them and from experience. An AI tasked with reading and summarizing the news on a particular sector might miss something important while the human analyst doing a similar task might pick up on a valuable piece of information which the AI failed to capture. This needs to be fed back into the system for it to learn and understand what the trader or analyst considers important or significant.” 

The paper referred to earlier by Wang, Choi, Xu and Yang details guidelines which can help direct this training. For example, they highlight the potential need for expertise around giving clear feedback and note that an organization needs to consider how to present what the model has learned and what feedback is actually needed. 

Making time for training 

Such developments highlight the training and time which financial services providers need to build into their plans to incorporate AI into their investment process. 

“Just like when you hire a junior associate, you need to allocate time from the people who can help train the AI program to do what you need it to. If you build that feedback loop within your data science team, your Ai shows up a little bit smarter the next day,” Ashley details. 

The time invested in training AI programs will be essentially won back in kind once the model is working as needed. It stands to reason that those who are just beginning to introduce the use of AI tools within their business need to plan to spend more time training the model. 

A paper by Accenture states: “Training different algorithms requires different amounts of data. Linear models can be trained reasonably well on a relatively small number of observations, and neural networks’ complexity means they need far more instances to learn from[2].” 

How much time it takes to train an AI program depends on the complexity of the model. Some can be trained in days or weeks, but then there needs to be a consistent process of fine-tuning and retraining, though to a smaller magnitude than the first training push. 

The OECD identifies the importance of this: “The robustness of AI systems can be reinforced by careful training, and retraining, of ML models with datasets large enough to capture non-linear relationships and tail events in the data (including synthetic ones). Ongoing monitoring, testing and validation of AI models throughout their lifecycles, and based on their intended purpose, is indispensable in order to identify and correct for ‘model drifts’ (concept drifts or data drifts), affecting the model’s predictive power.[3]” 

Managing expectations 

Ashley however points out that managers need to have reasonable expectations of these automated systems: “The main thrust of what they can do for an investor is manage data, prioritize that information and summarize it. This will allow traders or managers to be better informed than their competition; it will give them better context to help them make a decision. 

“However, there may be some unwillingness by hedge fund managers to allow AI to get too close to the decision-making process. There is a whole spectrum of how close you let AI come to your investment choices. At one end you have algorithmic traders and at the other are the discretionary investors who may worry AI will replace their skills.” 

“What if, instead of thinking of automation as the removal of human involvement from a task, we imagined it as the selective inclusion of human participation? The result would be a process that harnesses the efficiency of intelligent automation while remaining amenable to human feedback, all while retaining a greater sense of meaning,” muses Ge Wang, computer scientist and associate professor at Stanford University, in an article[4]. 

In an interview with Barry Hurewitz[5], Group Managing Director, Global Head of Smart Technologies & Advanced Analytics, Global Wealth Management, UBS, Chess Grandmaster Maurice Ashley observed: “Human to human interaction is something special to us and AI is not going to encroach on that. Social skills are lasting. Even if you are the manager of a tech company you still need those skills.” 

As processing power increases 10 times every five years according to  Moore’s Law, humans will be eclipsed by computers in many areas. Machines will bring lightning speed and accuracy to all manner of tasks. However, it would be a fallacy to assume that technology is making human effort redundant. It’s doubtful that computers will have fully mastered the fundamental, instinctive skills of intuition, judgment, and emotional intelligence that humans value by 2030. Over the next decade, partnering with machines will help humans transcend their limitations. 

This article should make clear that any apprehension of machines replacing humans can be quashed at this point in time. The feedback loop provided by people is one of the building blocks which can help AI integration to succeed. Therefore, managers should commit to training and maintaining the model and educating professionals on how it fits within the broader team. In fact, the BCG report quoted earlier expects that the top asset management firms ten years from now will be “bionic,” combining digital capabilities such as memory, speed, and stamina with human qualities such as empathy, creativity, and intuition. 

In collaboration with John Ashley, General Manager, Financial Services and Technology, NVIDIA.

This article is done in partnership with Dell Technologies & NVIDIA.


[1] https://arxiv.org/pdf/2103.04044.pdf 

[2] https://www.accenture.com/_acnmedia/pdf-110/accenture-financial-services-ai-neural-networks-pov.pdf 

[3] https://www.oecd.org/finance/financial-markets/Artificial-intelligence-machine-learning-big-data-in-finance.pdf 

[4] https://hai.stanford.edu/news/humans-loop-design-interactive-ai-systems 

[5] https://www.ubs.com/global/en/investment-bank/in-focus/2021/ai-gambit.html?campID=SOME-ADHOC-GLOBAL-ENG-LINKEDIN-IB-INSTREAM-AI-20210319-IMAGE-ALL-ORGANIC&sprinklrpostid=100001916449219

[6] https://www.dell.com/nl-nl/dt/solutions/high-performance-computing/index.htm#tab0=1&overlay=/collaterals/unauth/white-papers/products/ready-solutions/hpc-ai-algorithmic-trading-guide.pdf&tab1=0&accordion0 

Like this article? Sign up to our free newsletter

Most Popular

Further Reading

Featured