@JoakimPrestmo: Last day of deep learning in finance summit #reworkFIN This'll be a great day!

‘Nearly 73% of everyday financial trading is executed by machines. Every major financial firm is investing in algorithmic trading because the level and volume of trade carried out by these machines is out of human bounds to process and execute...these machines take into account the past historical financial data available [and] maximize their returns.’ iamWire

Back for day 2 of the Deep Learning in Finance Summit, we heard from the startups using AI and deep learning (DL) to drive their profit margins and make smarter, machine driven decisions.

Jakob Aungiers from HSBC and our compare for today kicked off this morning’s discussions by introducing four FinTech startups who are disrupting the industry with their implementation of AI and DL.

‘With cyber attacks getting more fierce and dangerous, the basic rule stands: Do not fall for the temptation of opening attachments or links. We see advice like this all the time. But how do we really avoid cyber attacks?'

Aeneas Wiener, CTO and Co-Founder of Cytora began by discussing how insurers ‘rely on comprehensive historical claims data sets to price risk accurately, but claims data does not exist for new and emerging risks such as cyber attacks’. Simple advice as above isn’t enough to protect us,

@richcasselberry: 90% of physical world events will be recorded in unstructured by 2020 @ValaAfshar Almost sounds like you #reworkfin #techatliberty @cytora

Wiener explained how they are using machine learning (ML) to turn unconstructed data into commercial insight to help insurers understand risk in an empirical way. They have ‘built a risk engine that structures multilingual unstructured text using natural language processing, and use AI machine learning algorithms to train these models to structure the mountain of unstructured data’. These include neural network based sentence multi-class classification and unsupervised topic clustering, and can be applied to create synthetic loss event data sets from openly available data. For example, in the case of a restaurant Cytora collect data such as opening hours, website, whether or not they do takeaway, locations etc. and they have found that, using this example, ‘a restaurant in a city centre that is open late, serves takeaway but has no website has 30% more risk than other restaurants.’ The algorithm they have created to read and aggregate these formulas across a variety of data takes roughly a day, whereas it would take a human 5 years to read the same data.

We also heard about how this approach can generate loss frequency models suitable for pricing risks like automotive recalls, factory fires and much more. Insurers are no longer limited to their own data claims, and ‘high-resolution predictive rating factors derived from unstructured web data allows these projections to be much more accurate.’

@cytora: Thanks for having us @reworkfinance! Fantastic to share the stage with an excellent lineup of speakers #DeepLearning in #finance #reworkFIN

Deep learning faces 3 key challenges, and Giacomo Mariotti from Alpha-i outlined these:

‘Firstly, it’s a huge model that needs a lot of training and requires the data to be calibrated. Secondly, the model is overconfident and doesn’t know what it hasn’t been taught, and thirdly even state of the art AI requires a lot of human interaction to be able to calibrate the data.’

Alpha-i  are overcoming these obstacles by the implementation of Bayesian methods combined with DL.

With a traditional neural network, inconsistencies appear where ‘the model doesn’t really know what’s happening ,and produces inaccurate data’. This is where they Bayesian method comes in to assist the neural network, bringing ‘extra information that helps us question what the model is telling us.’

The Alpha-i DL network is able to make forecasts from a time series as well as associating each prediction with a confidence level, which is derived from the information about the model and the data available. He continued the discussion by telling us how they are ‘leveraging on the power of DL methodologies with the aim of delivering accurate time series forecast with their uncertainties’.

One of the key aspect of this Bayesian deep learning methodology is its aversion to over-fitting obtained thanks to the robust probabilistic inference framework. Bayesian inference methodologies can significantly boost the online performance of Alpha-i’s machinery and they are currently working to optimise and improve this.

Learn more: Yarin Gal - What My Deep Model Doesn't Know

ML can be extremely complex, but it’s capabilities are undeniable. ForecastThis want to make ML accessible to the investment management community:

‘We don’t want people to have to have a PhD to be able to implement ML - it’s all about what you can do with the tech.’

Justin Washtell-Blaise, CEO & Co-Founder explained how the investment management industry has been cautious about embracing these new technologies due to these barriers. To allow the industry to implement deep learning techniques, ‘we have to make it accessible and relevant - we don’t want to dumb it down and it needs to be robust enough to be used in a turnkey fashion without compromise.’

DL and related techniques are helping ForecastThis to close this gap and they are working with a functional API, as well as building on a graphical user interface to make this more accessible. ‘ML and DLis still a bleeding edge technology and we need to think carefully about our relationship with it.’

ForecastThis are working with both humans and computers together to produce optimum results. For example, whilst ‘computer power has been getting better and better for the past 20 years, but the best chess players today are still human-computer teams.’ He emphasised that we shouldn’t put too much trust in computers due to DL still being such an unknown area. However, this collaboration does not mean ‘having a human gatekeeper overriding computer decisions’, but using human logic to assist the features. He expanded on understanding the uncertainties and using visualisation techniques to leverage the brain’s inherent computing capabilities and comprehension.

In the Q&A, our compare Jakob asked whether working together with DL will lead to job cuts? Washtell-Blaise answered ‘I hope not, but the evidence from chess, not that I’m saying chess and investiment management are the same thing, but it’s commons sense to have humans in the loop - our minds work differently to machines, I can’t say how it’s going to work’.

@reworkfinance: "Chess is a problem that we thought computers had solved" Justin Washtell-Blaise from @forecastthis #reworkFIN

What usually springs to mind when we think of investment banking? Daniele Grassi, CEO & CTO of Axyon AI summed it up pretty well saying ‘we think of Wall Street and people in nice suits’, and he went on to tell us how this isn’t the whole story. Investment banking represents a high-margin finance field, and he expanded on the notions from this morning that in this industry AI and DL are still under-exploited.

In investment banking, the data used comes from 2 main sources: ‘data providers which are available to subscribers, and proprietary data which is keep internal to the financial institution.’ The data is spread across several databases and applications from data providers which is often unlabelled and disjointed. However, for DL to be successful, large amounts of data are required and this is where investment banking has a huge advantage due to the depth of the data.

He expanded on the barriers of domain knowledge, data quantity (it’s often spread over a variety of databases), and privacy and licensing terms, and explained how they are using DL to overcome these issues. ‘The most crucial challenge of all is moving from accuracy to value and defining what it means for your model to be successful, and have a fully functional product to demonstrate to your client.’ Axyon’s DL solution provides accurate predictive models, and helps corporate clients optimise their processes as well as reduce risks and increase revenue.

Running alongside this summit, we’ve also had our Deep Learning in Retail and Advertising Summit, and you can view the day 1 highlights here.

Couldn’t make it to London?

Register for on-demand post-event access to receive all the slides, presentation videos and interviews from the summit, or check out our calendar of upcoming events here.

We will also be back in London in September as part of our Deep Learning Summit tour. Find out more about the summits here: