Rogue Trading
Rogue trading is an issue that has plagued multinational banks since their inception. Defined as a trader who acts independently of others. The banking industry as a whole has lost upwards of £10 billion due to rogue trading, the most recent of which was Jerome Kerviel who lost £3.7 billion in 2008. In some cases, these losses have brought down whole banks as is the case with Barings Bank and Nick Leeson.

Examples
Societe Generale – £3.7 billion
Barings Bank – £827 million
Allied Irish bank – £697 million
Daiwa Bank – $1.1 billion
Sumitomo Corporation – $2.6 billion

How it Happens
Rogue trading occurs when a trader covers up trades by booking a trade that hedges directional risk. Typically, the trading will “book” these hypothetical hedge trades so that they can cover up large positions. If these positions go south, the bank can stand to lose a lot of money. The compensation structure means that there is an unlimited upside (larger bonus) for the trader but limited downside (getting fired).

How it Can Be Prevented
Policies and strict governance can prevent rogue trading but another more direct way is stricter controls on the technologies used for booking these trades. Inefficiencies and inconsistencies in databases used within banks may allow traders to hide these trades from the central risk book. DataSpartan worked on a Blockchain proof of concept platform that addressed these database inconsistencies.

How Does Blockchain Help
A Blockchain type solution allows different departments within the bank to have a consistent view of orders – this is because it is impossible to delete or change any of the executed trade orders without other departments agreeing. This provides transparency about trade execution and minimises the chance of rogue trading.

Our Results
The proof of concept is now being fully integrated into a subset of the servers in the multinational bank for live testing. APIs have also been built to allow developers to interface with Blockchain functions.

Client Classification
For business to business enterprises, client classification focuses on putting a customer into a specific category or bucket. In this specific scenario, DataSpartan was working with a crowdsourcing startup who wanted to identify whether or not a business on their platform was likely to “hit” their crowdfunded target amount raised. The client also wanted us to inform them of which variables were most important in terms of increasing the likelihood of a raise.

Problem
A large proportion of the onboarding process was done manually and it was extremely time consuming for the startup to fully onboard each client only to have them not complete a successful raise. Standardized data was collected about the size of the company, the number of followers they had in social media as well as financial performance, their team, industry and competition as well as exit potential. Currently, 56% of the companies that were going through the funnel had a successful raise.

Solution
The first part of the problem was identifying which variables were the most likely to influence the outcome. Historical data was extracted, cleaned and standardised to allow this analysis and three variables were singled out which determined the probability of a successful raise. A supervised learning model was then created which could classify these leads based on their likelihood of conversion based on their data. Python was used for this portion of the data analysis and libraries such as NumPy and SciPy were leveraged to ensure rapid computation and fast model prototyping.

Our Results
The current research and results have been incorporated back into their lead pipeline in the form of a qualifying questionnaire which filters out unlikely leads and pre categorises the clients. This has resulted in a 4% decrease in the number of man hours used as well as a 6% increase in the number of companies that successfully raise. Further work is being done to automate some of the manual processes such as KYC (Know-Your-Client) document verification to allow the client to efficiently scale up its business. Work is also being done to improve the data pipeline and to ensure that the data collected is being stored in the appropriate format for analysis.

Price Prediction
Since the start of the electronic market, price prediction has always been a hot topic. With the advent of machine learning, the field is becoming more and more contested with mathematical methods leveraging these new tools. Stories from hedge funds such as Renaissance Capital have inspired financiers around the world to look towards mathematics as a means of generating alpha.

Data
Two quants from DataSpartan were assigned to work on and analyse medium frequency and high frequency data respectively. They were asked to specifically investigate bond and option prices over a one hour period and a millisecond period to determine if there are any patterns that could be discerned from the data. The client had a proprietary data feed of tick data for each symbol which allowed the quants to conduct in depth research.

Approach
The client had a preexisting framework written in Java which could instantaneously switch between backtest and live. Ideas were gained mainly by reading research papers in the physics, biology and mathematics fields – the ideas were then explored and implemented on time series data. Strategies had to be run for a minimum amount of time and deemed profitable in a backtest before being put onto live servers. Concepts which had predictive characteristics such as entropy were explored (entropy has been used to predict the probability of cardiac arrest during open heart surgery by investigating the differences in systolic and diastolic pressure).

Our Results
After four months of testing, a strategy was found that yielded a back tested Sharpe ratio (a measure of the reward over risk) of 1.5 inclusive of commissions. This strategy is now being run live on the platform and is currently generating profits for the client. Work is currently being done on the strategy to check if it works on other instruments in different asset classes. Due to the success of this strategy, the client has has DataSpartan to do some further work investigating strategies for portfolio allocation and diversification on a macroeconomic level. They want to investigate the automation of Warren Buffet’s trading style by analysing 10-K reports using machine learning and determining the optimal allocation of capital on an industry basis.

Purpose
Our client lacked the in-house talent required to develop a generic trading platform for back testing strategies. DataSpartan was brought in to resource and manage the project.

Approach
The backtesting system was built in Java and used Spark Streaming, Kafka and HDFS to build a continuous exchange-transform-load (ETL) pipeline for gathering the data feeds of stock prices and in particular streaming social media data in order to produce short term predictions of the stock price.

Our Results
The system has been fully built and currently DataSpartan quants are building on top of the system in order to identify profitable strategies in both the high frequency and medium frequency space for market making. Currently, a market-impact tested strategy of Sharpe ratio 1 has been found and is being put into production.