In depth: Could speech analytics prevent the next financial crash?

[sincity]

Introduction

The public furore over dodgy bankers may have subsided a little, but since the 2008 financial crisis we have seen 2013's rigging of Libor and the recent fixing of the foreign exchanges market.

Fines and compensation payments (read: PPI) have cost British banks £38.5 billion (around $60 billion, AU$73 billion) this century, and some think it's the banking culture that's to blame. However, there could be a way to change this bad behaviour, and TechRadar Pro was given a demo of this tech. It's a new speech recognition and analytics engine that records and scrutinises every single sentence – plus emails, chat rooms, IM and texts – to identify fraud as it's happening. This fix could be history.

Why do we need speech analytics?

Although banks need to limit their risk to fines and compensation, the new market for speech analytics is opening because of the Dodds-Frank Act in the US, which was created after the collapse of the Lehman Brothers bank in 2008.

It puts an onus on investment banks, in particular, to be able to reconstruct derivative and swap trades within just 24-72 hours. "The Act actually says the data should be 'readily available', which could be as short a time as 14 hours," says Juan Manual 'Juanmo' Soto, an industrial engineer and CEO of Spanish linguistics company Fonetic who specialises in speech analytics. "Every big bank in the world is affected."

His solution is the Fonetic Linguistics Voice Platform, which uses a unique algorithm that records and analyses speech. "It's about managing risk for banks," says Karen Winter, Sales & Marketing Director (EMEA) at Fonetic, who began her career in the banking sector working for Morgan Stanley, and later worked as a foreign exchange trader for ABN AMRO Bank. "The nature of trading is to push the boundaries, and sometimes bankers go over them."

Santander currently uses the Fonetic Linguistics Voice Platform in seven trading rooms across three continents. Other customers include another Spanish bank BBVA.

How does it work?

"We use the Nuance speech recognition engine, and Genesys SpeechMinor as the platform for handling the administration of the calls," says Juanmo, though the crucial speech analysis algorithm is unique to Fonetic.

The Linguistics Voice Platform does two things. Firstly, it allows a trade to be reconstructed. "The trade isn't closed so there is no reference number," says Winter about the detective work involved in tracking the history of a fraudulent trade. "You have to use metadata like the date and the content of the conversation to detect what the underlying asset is, and who was involved."

The second use of the tech is in Trading Communications Surveillance, the monitoring in real-time of voice and, crucially, text sources also – including emails, chat rooms, and other written documents on a corporate network – for fraud and bad practices. Anything that the bank wants to detect, it can generate alerts for in near real-time or compile a list of suspicious conversations for auditors to look at later.

Can it understand different languages?

Yes – and speaking Klingon doesn't fool it since the software immediately flags up anything unusual. Nuance's speech recognition engine can handle 84 different languages, and many variations within them, and the Linguistics Voice Platform software can listen out for all of them.

"It's working in five languages for some banks," says Juanmo. "The Nuance engine even includes variations – it can understand US English, Australian English, UK English and even Singapore and South African English, with four or five versions of Spanish also," he adds.

Big data and catching fraudsters

Dealing with big data

Traders' calls have been taped since the 1970s, but they've never been analysed until now. "The average trader will generate 2,000 hours of recordings every year," says Winter. "Multiply that by 500 on a small trading floor then multiply that by five trading houses … that's 5.6 million hours of calls per year!"

Since the Dodd-Frank Act defines that the lifetime of a swap is 35 years, it's a lot of data to sort and archive. "A proper trade reconstruction is impossible," says Winter. "The banks have not got a chance."

Dealing with city-speak and slang

Before computers, traders dealt with each other in coffee houses and other public places, so they developed their own language to keep the details private. It's a tradition that's stuck, but it's about to become redundant. Take this example of a concluded deal picked-out by the Linguistics Voice Platform: "Hi how are you, mine." Now that's subtle, but easy to spot if you're looking for it.

"Traders shortcut a lot – time is money – and prices change by the second," says Winter, who's steeped in bankers' lingo. "We put context into language, so if something looks out of context, it's flagged and categorised," she says.

Can it catch fraudulent bankers?

Absolutely. Here's an example of how the tech operated during the period of Libor fixing – we've capitalised the examples of word strings or phrases the software picked out:

"This is the way you PULL OFF DEALS LIKE THIS CHICKEN, DON'T TALK ABOUT IT TOO MUCH, two months of preparation ... THE TRICK IS you MUST NOT DO THIS ALONE … THIS IS BETWEEN YOU AND ME but REALLY DON'T TELL ANYBODY."

The rest of that conversation was punctuated with promises of Bollinger champagne alongside odd sounding compliments such as 'big boy', which were all flagged by the tech.

Another more subtle – but detectable – Libor-related conversation went like this:

Senior euro swaps trader: "Hi, is it too late to ask for a low 3m?"

Euribor submitter: "Just about to put them in ..... so no."

However, whether this kind of technology can prevent the next financial crash entirely depends on the reason for a fall in the markets. That could be fraudulent behaviour, though that wasn't the issue last time around in 2008.

"The banking crisis was to do with liquidity in the market, and that's what triggered Dodd-Frank to make the market more competitive and liquid," says Winter. "Whether we could have helped, who knows, but we do know that with our solution trader behaviour would probably have been different."

The recent Forex foreign exchange rate fixing scandal, however, could have been prevented by the Linguistics Voice Platform because it works across all sources of communication, both written and spoken.

"That's just a bunch of guys who got together probably in a chat room," says Winter. "Markets move on events – they need an election, an event to move the market – so these guys got together to put their money together to move the market between them. They've been fixing their position prior to their market move then selling out at the top of that position to maximise profits – it's highly irregular."

The software spots that kind of behaviour easily because it looks at trading patterns, and because criminal activity tends to happen in short periods of the day. "Once you are suspicious, you can track everything," adds Juanmo.

Does it work in other industries?

Anywhere that has a high volume of calls, emails and other types of communication could benefit from speech analytics.

"The software is found in call centres, in the telecoms industry, energy, insurance – any company handling big volumes of calls which they want to analyse," says Juanmo, picking out three examples of how speech analytics can produce data to improve efficiency: the churn in the telecoms sector, calculating customer satisfaction in call centres, and processing insurance claims. The platform is already used by Vodafone, Telefónica and Direct Line.

Users' 'voice-prints' will soon be integrated to increase security, while sentiment analysis is now possible, and already used by Fonetic in the call centre-centric versions of its Linguistics Voice Platform to detect underlying feelings of anger or frustration in customers' voices by studying intonation. That doesn't apply to banking, which is largely about facts. That's not going to change, but banking culture is about to be shaken up – and shaken down – by the sheer thoroughness of speech analytics.