OUR RESEARCH

We advance the field of A.I. research by tackling practical challenges.

Our poster presentations.

For our students working on short term projects (between 4 weeks up to 1 year), we often encourage them to make poster presentations as a means to explain their work and contributions to others such as business stakeholders. This further develops student confidence and communication skills. Here are a few examples:

HOW DO ECONOMIC RELEASES AFFECT PRICE FX PRICE MOVEMENT?

In these posters (here and here) our students looked at the price movement over 10 years of historic Foreign Exchange (FX) prices and Economic Release Calendar events. We showed that the severity of the releases and the interpretation of the figures released gave good forecasts of jumps in volatility and price direction.

WHICH CLIENTS TAKE THE BIGGEST RISK IN FX?

Using 3 years’ worth of client trading activity over a space of 5,000 clients from a global brokerage, our students investigated the relationships between:
The clients background: such as their age, gender, location, income etc
How they tended to trade: such as what symbols they traded, their risk appetite with Take Profit / Stop Loss, averaging time holding positions etc.
You can view this poster here

Our peer-reviewed publications.

For our students working on longer term projects (greater than 1 year, e.g. our PhD students) we like to encourage them to submit aspects of their work for publication either at a peer-reviewed academic conference and/or within a journal. Responding to feedback from peer reviewers is an essential part of the learning process and referee comments in addition to conference discussions can often inspire new lines of enquiry within the project as a whole. Students undertaking a long-term project have time to explore A.I. topics in greater depth, increasing their chances of making novel contributions to the field. What is more, attending conferences offers an excellent opportunity to stay up to date with the fast-paced field of A.I. research.

STRUCTURING TIME SERIES DATA TO GAIN INSIGHTS INTO AGENT BEHAVIOUR

Here we introduced a novel data wrangling approach – Data Aggregation Partition Reduction Algorithm (DAPRA) to convert irregularly sampled time series data into regularly sampled for effective visualisation. In the study we investigated two distinct problem domains that of financial and travel datasets.

Al-baghdadi, N., Wisniewski, W., Lindsay, D., Lindsay, S., Kalnishkan, Y., Watkins, C. 2019, ‘Structuring Time Series Data to Gain Insight into Agent Behaviour’. Proceedings of the 2019 IEEE International Conference on Big Data (Big Data), Los Angeles, 9 – 12 December, 2019, pp. 5480 – 5490.

PRACTICAL INVESTMENT WITH THE LONG SHORT GAME

In this research we extended a well-studied online learning technique called the Aggregating Algorithm and showed how it can be developed and improved to better manage portfolios of experts with diverse long and short positions across many assets. We carried out our experiments using a real world data set comprising a large portfolio of FX basket trading activity.

Al-baghdadi, N., Kalnishkan, Y., Lindsay, D., Lindsay, S. 2020, ‘Practical Investment with the Long-Short Game’. The 9th Symposium on Conformal and Probabilistic Prediction with Applications, Verona, 9 – 11 September, 2020, PMLR 128:209 – 228.

APPLICATION OF CONFORMAL PREDICTION INTERVAL ESTIMATIONS TO MARKET MAKERS’ NET POSITIONS

Using a real world time-series data set that spanned 5 years’ worth of net positions from a financial brokerage, we investigated a variety of implementations of Conformal Predictors (CP) on top of standard machine learning techniques (LSTM’s, K-Nearest Neighbours etc), as well as standard techniques such as quantile regression and simple moving averages, to show that CP gave both practically useful and statistically valid confidence bounds.

Wisniewski, W., Lindsay, D., Lindsay, S. 2020, ‘Application of conformal prediction interval estimations to market makers’ net positions’. The 9th Symposium on Conformal and Probabilistic Prediction with Applications, Verona, 9 – 11 September, 2020, PMLR 128:285 – 301.

TEMPORAL DISTRIBUTION OF CLUSTERS OF INVESTORS AND THEIR APPLICATION IN PREDICTION WITH EXPERT ADVICE

Using a real world data set capturing the trades of over 10,000 traders across a 3 year period, we investigated how trader activity can be clustered and classified through time. We explored the use of several clustering techniques and demonstrated how they can compress the data set, removing redundancy and helping improve online portfolio allocation techniques such as the Aggregating Algorithm.

Wisniewski, W., Lindsay, D., Lindsay, S., Kalnishkan, Y. 2022 (submitted for review), ‘Temporal distribution of clusters of investors and their application in prediction with Expert Advice’. – submitted to EPJ Data Science

ONLINE PORTFOLIO HEDGING WITH THE WEAK AGGREGATING ALGORITHM

Working with a real world time-series data set tracking a market maker’s net position over 3 major FX symbols, we demonstrated how the Weak Aggregating Algorithm (WAA) could be used to combine the trading decisions of a set of different simple “Cylinder” hedging models to improve the profitability and the stability of the Profit and Loss (PnL).

Al-Baghdadi, N., Kalnishkan, Y., Lindsay, D., & Lindsay, S. 2022, ’Online Portfolio Hedging with the Weak Aggregating Algorithm’. The 11th Symposium on Conformal and Probabilistic Prediction with Applications, Brighton, UK, 24 – 26 August, 2022, PMLR 179: 149 – 168.

Our publications before AlgoLabs.

Though the publications below are over 15 years old, we’ve included them here to show some of the early research that AlgoLabs’ founders Siân and David worked on together when they themselves were students.

IMPROVING THE RELIABILITY OF DECISION TREE AND NAÏVE BAYES LEARNERS

Decision tree and Naive Bayes learners are known to produce unreliable probability forecasts. We used simple binning and Laplace transform techniques to improve their reliability and compared their effectiveness with that of the Venn probability machine (VPM) meta-learner. The VPM outperformed the simple techniques to improve reliability, although at the cost of increased computational intensity and slight increase in error rate.

Lindsay, D. and Cox, S. 2004, ‘Improving the reliability of decision tree and naive Bayes learners’, Fourth IEEE International Conference on Data Mining, Brighton, UK, 01 – 04 November, 2004, ICDM’04: 459 – 462.

EFFECTIVE CONFIDENCE PREDICTION USING PROBABILITY FORECASTERS

We saw how machine learning approaches could be extended to make valid confidence prediction bounds toward practical benefits in medical diagnostics.

Lindsay, D.G., Cox, S. 2005, ‘Effective Confidence Region Prediction Using Probability Forecasters’, International Conference on Artificial Intelligence in Medicine, Aberdeen, UK, 23 – 27 July, 2005, AIME’05: 494 – 503.

EFFECTIVE PROBABILITY FORECASTING FOR TIME SERIES DATA USING STANDARD MACHINE LEARNING TECHNIQUES

Machine Learning (ML) techniques typically assume data is i.i.d. meaning each data point is drawn independently from an identical distribution. Time series data typically violates both these assumptions, yet our study showed across a variety of real world data sets that standard ML techniques can still produce valid probability forecasts.

Lindsay, D., Cox, S. 2005, ‘Effective Probability Forecasting for Time Series Data Using Standard Machine Learning Techniques’, Third International Conferences on Advances in Pattern Recognition, Bath, UK, 22 – 25 August, 2005, ICAPR’05: 35 – 44.

LEARNING FROM STRING SEQUENCES

In my second-year PhD technical report (April 2004) “Learning from String Sequences,” I applied the Universal Similarity Metric (USM) as an alternative distance metric in a K-Nearest Neighbours (K-NN) learner for effective pattern recognition of variable length sequence data. By comparing USM with the string-to-word vector approach across spam email filtering and protein subcellular localization datasets, I demonstrated that the USM-based K-NN learner outperformed traditional techniques, yielding higher classification accuracy and reliable probability forecasts.

MONDRIAN CONFIDENCE MACHINE

In the first year of his PhD, David helped his supervisor Volodya Vovk with research on the Mondrian Confidence Machine (MCM), an advanced online prediction algorithm that outperforms previous methods by addressing diverse prediction issues and maintaining high calibration accuracy across different types.

TRANSDUCTIVE CONFIDENCE MACHINES AND THEIR APPLICATION TO MEDICAL DATASETS

In my third-year undergraduate dissertation (Mar 2002), I explored machine learning techniques for medical datasets, focusing on the development and testing of the Transductive Confidence Machine Nearest Neighbours (TCMNN) algorithm. The project also investigated the impact of various parameters, Minkowski metrics, and polynomial kernels, and compared the TCMNN algorithm with SVM and neural network implementations.