Artificial intelligence is a very hot topic
Contrary to what people might think, artificial intelligence (AI) is hardly a new topic. It has been around since 1956 when the seminal summer workshop was organized at Dartmouth College, New Hampshire, US.
For years, artificial intelligence remained a subject of scholarly study or an inspiration for science-fiction writers. However, there has been a significant acceleration in recent years. AI has started to be implemented for real-world applications, including in business contexts. It’s happening for three reasons:
Data is available: our digital world is producing at an ever-increasing rate an incredible amount of both structured (databases) and unstructured (files, images, videos) data. Data is the “new oil” that intelligent algorithms consume: the more data is given in input, the more accurate the prediction output is.
The computing power is available: thanks to Moore’s law, in effect for the last 50 years, processors have become efficient enough to analyze the data at a reasonable cost in a reasonable amount of time.
Breakthroughs in algorithm efficiency: complex algorithms such as speech recognition have improved over the years, finally reaching the accuracy level of humans in 2017.
Consequently, venture-capital (VC) investments in artificial-intelligence startups have increased sharply in recent years, from less than $500 million in 2007 to more than $6 billion for the first seven months of 2017, according to Venture Scanner. At the same time, the main technology companies have been on a buying spree. For instance, Google has bought 12 AI companies since 2012.
Thanks to this interest and flow of money, there has been an explosion of new entrants aiming to apply artificial intelligence in different areas of finance, more than 100 startups, according to CB Insights.
What intelligent algorithms can do
There are three types of machine learning:
Unsupervised learning: using statistical tools for data clustering, to find “hidden” patterns, without any external feedback (e.g., relevant customer segmentation),
Supervised learning: a machine is trained for a specific classification task using labeled data and direct feedback (e.g., credit worthiness of customers),
Reinforced learning: algorithms learn to react to an environment by repeating strategies over and over while maximizing rewards (e.g., adjustment of a sale offer based on acceptance/rejection rates).
Machine-learning algorithms are typically used for voice/language recognition and generation (e.g., chatbots), image recognition (e.g., self-driving cars) or to solve specific business problems.
1/ Investing – asset management: algorithms can be used to search for correlations between world events and their impacts on asset prices, or to learn from publicly available social-media streams to anticipate markets’ movements (e.g., Kensho, Dataminr).
2/ Credit scoring – underwriting: machine learning can help lenders make more accurate credit-underwriting decisions, or advanced computer vision can be used with geospatial and aerial imagery for insurance/property underwriting (e.g., ZestFinance, Cape Analytics).
3/ Regulatory compliance – fraud detection: different channels and types of data can be analyzed with advanced pattern-matching analytics to detect fraudulent activity (e.g., Digital Reasoning, Actimize).
Today, a typical anti-money-laundering process will perform an automated scan of incoming and outgoing payments based on predefined rules (country of origin/destination, name of the customer, etc.). Current systems generate a lot of false positives that are reviewed one by one by middle-office operators and/or compliance officers. Machine learning can be used to identify users to add to the whitelist, identify patterns to be added to the rule engine and ultimately reduce the number of false positives, saving costs while increasing the quality of the screening process.
4/ Market research – reporting: intelligent agents can curate and semantically index the financial-markets research content, and automate the writing of reports, personalized websites, emails, articles and more with natural-language-generation software (e.g., AlphaSense, Narrative Science).
5/ Customer support – assistants: intelligent agents can analyze incoming messages, route cases, provide customer-services agents with accurate suggestions, or help optimize personal-finance management (e.g., DigitalGenius, Pefin).
The challenges of artificial intelligence
Technology “evangelists” excel at creating the buzz around artificial intelligence by focusing on its promises. In the real world, however, reaping the benefits from intelligent algorithms can be very challenging.
1/ Data quality:
The prediction power of an algorithm is highly dependent on the quality of the data fed as input. Even in quality sources, biases can be hidden in the data. The time and effort required to gather and prepare an appropriate set of data should not be underestimated. For the nascent self-driving automotive industry, for instance, most of the effort is spent on labelling hours of videos. This need has led to the creation of an entire offshore industry for video labelling.
In the financial industry, the reconciliation of the data from front to back is already problematic, and data referentials are often plagued with quality issues. Having a data-quality program in place is a prerequisite to any large-scale artificial-intelligence initiative.
2/ Black-box effect:
The results of intelligent algorithms are opaque and not verifiable. They deliver statistical truths, meaning that they can be wrong on individual cases. The results could have a hidden bias difficult to identify. The diagnosing and correcting of those algorithms is very complex.
The fact that there is no explanation as to why the algorithm provided a positive or negative answer to a specific question can be disturbing for a banker’s rational mind. This is often a blocking point for the use of AI in trading.
3/ Narrow focus:
By design, intelligent algorithms are good at solving specific problems and cannot deviate from what they were designed for. An algorithm trained to detect suspicious payments would not be able to detect any other suspicious activity related to trading, for instance.
In addition, algorithms are purely rational and lack essential factors such as emotional intelligence and the ability to contextualize information, unlike human beings. That’s why banking chatbots often disappoint: they are “smart” but lack empathy.
The use of intelligent machines represents a challenge in terms of liability: who/what shall be responsible in case something goes wrong? Financial institutions are reluctant to give machines full autonomy because their behavior is not fully foreseeable. They tend to keep a human supervisor to validate the machine’s decisions for critical activities such as releasing/blocking payments or validating trades, partially defeating the purpose of using a machine in the first place. Current compliance and operational security standards are quite strict; I anticipate that they will loosen over time when the technology matures.
The global risks for incumbents
Until recently, large financial institutions could fend off competition thanks to the scale of their operations and their information advantage. They could run expensive datacenters and hire large research teams. It was impossible for startups to compete. Nowadays, data scientists fresh from MIT (Massachusetts Institute of Technology) or Harvard can literally launch a fund using advanced machine-learning algorithms by leveraging cloud-computing services. Information is still money, but information is now more and more distributed, accessible and exploitable by small actors.
We might soon witness a role-reversal situation. Before financial institutions could hire technology experts to support their growth; now we see the Googles and Amazons of the world starting to hire business experts (traders, underwriters, etc.) and compete directly against established actors!
Regulation, while being a burden on the operations of incumbents, is still protecting the industry from a quick disruption. But for how long? Can financial institutions put up with just buying young competitors and integrating their products into their own services?
Where to start with artificial intelligence
Because the concept of “artificial intelligence” is very broad and because its application to finance is recent, financial institutions often struggle with how to structure their innovation approach to machine learning:
- How to select the right use-case for experimentation?
- How to scale successful proofs of concept?
- How to integrate the new tools within the IT (information technology) legacy?
- How to develop and organize/govern an internal center of expertise?
It can be tricky to navigate a maturing market. Recently one of our clients wanted to select a tool for a proof of concept and received bids from $20,000 to $1 million!
When structuring your approach, keep in mind that:
Innovation is about business innovation—technology is only an enabler. Always start from business needs and pain points and avoid the “technology looking for a solution” conundrum.
Innovation is not necessarily “disruptive”—define a balanced portfolio of initiatives from incremental improvements to more transformative concepts.
Innovation can be sourced internally and externally—the key is to find the right balance.
Idea generation and creative brainstorming are necessary but not sufficient—to succeed, innovation should be considered as a global system, from strategy, governance, procedures, to sourcing and culture.
Artificial intelligence is still at an early stage. Because of its inherent challenges, the first implementations usually don’t bring huge benefits. However, it must not be ignored. It will profoundly change financial services. It is never too late to start the journey. Start now!