Visualisation de l'information appliquée à l'analyse et à l'attribution de performances financières

Abstract

Croesus Finansoft has been developing portfolio management software for brokerage firms and independent advisors for the past 28 years. Their solution is currently used by most large Canadian brokerage firms, including CIBC, National Bank, Desjardins Wealth Management and TD. Consequently, the software must manage large data tables, often comprising more than a billion rows. Data size is a challenge for Croesus at every level. The software must give insights on investor’s portfolios, guide advisors on new investment possibilities, follow investment performances, etc. Even with current technological advances, some issues remain difficult to solve, mostly because of the amount of data involved. High level portfolio performance analysis is especially a challenge. Performance analysis requires more than simply comparing returns at different points in time. It is a complex process requiring data correlation from multiple sources to obtain a coherent portrait of the situation. Investment performances are always evaluated against a benchmark, such as a market index. Performance attribution tries to explain where this differential return comes from. Is it because investors picked better stocks than the one used in the benchmark ? Is it because more money was invested in long term bonds compared to the index, limiting the risk exposure ? Croesus’ software tools allow a user to easily measure performance for a single or a small subset of portfolios. Performing such analysis on a larger scale, such as for all clients in an office, would be much more complex. In addition, their software does not perform performance attribution so far. For Croesus’ managers, these features are real challenges, mostly because of the amount of data implied. How should we display this information to the expert without overwhelming him ? How can we easily identify data problems, global trends or other performance issues in order to act before it is too late ? Information visualization enables the leveraging of human image processing power to interpret data much faster than any numeric and textual ways. Visualization can augment human capacities in order to keep the user in the decision loop, and not replace him completely like most automated decision tools. Even though visualization has been an active research field of interest for many years, only a few contributions have been published related to finance and performance analysis. This thesis explores different visualization techniques in order to simplify financial performance analysis, related to Croesus portfolio management requirements, and presents results from three different project related to performance analysis. The first projects presents a novel interaction technique allowing the simplification of performance analysis on a simple line chart. Whether the chart is comparing returns from several hundreds of portfolios or stocks from different industries, such line charts are easily overwhelmed with the amount of data, mostly because of occlusion. Our proposed tool, VectorLens, allows extraction of interesting elements using advanced selection techniques. Our main contribution is related to angular selection. Since the charts present returns against time, the slope encodes important information. VectorLens takes advantage of this situation and allows the selection, using a single drag, all elements whose slopes fit within a dynamically adjustable range. VectorLens also integrates other selection techniques, such as brushing and categorical selection. It is also possible to combine multiple VectorLens lenses to create more complex queries. Angular selection was compared experimentally against state of the art techniques and was proven more or equally efficient in most cases, and was preferred overall by users over the other more conventional techniques. Our second project proposes a new visualization technique for effectively separating data layers on a line chart. This technique is useful for comparing stocks from different industries on the same chart, or for comparing multiple portfolios from different advisors, for example. Instead of using only color to separate layers, this technique introduces in-place layer compression at each tick, avoiding the usual occlusion caused by overlapping layers. Many variations were created using this concept and were compared against state of the art techniques in a user study. Compression was shown to be more efficient, for example, when comparing values between different layers on a single day. Finally, our last project addresses the large scale performance attribution issue. Two novel visualization techniques based on the ternary plot were proposed in order to present both differential returns and attribution effects on a single chart. A complete dashboard integrating these new charts was also created and was evaluated by four domain experts in a case study using real data. Results showed that the proposed tools enable easy analysis of large data sets at different levels. The proposed tools also clearly display performance deviations and their root causes. They also clearly illustrate advisors’ overall strategies and outlier accounts.

Date	14 Dec 2015
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Michael John McGuffin (Supervisor)