Alexandre Andreani Explains His Short-Term Intraday Quantitative Trading Strategy
Alexandre Andreani is founder of Andreani & Associ©es, a quantitative investing laboratory of emerging hedge funds strategies constituted by a team of scientists, econometricians and mathematicians, all PhDs and post-doc from top universities in Europe and in USA with over 10 years of experience in research.
Alexandre Andreani served as general partner of a quantitative investment advisor based in Geneva, and portfolio manager and principal advisor of several hedge funds, for which he developed all of the proprietary trading algorithm and investment strategies, working with family offices and institutional investors. He has recruited, trained, and managed teams of quantitative analysts in Europe for both money management and research firms.
Mr. Andreani has been working for Quantitative Investment Research in Geneva, as head of the investment committee, CEO/CIO of the company and Global Head of Research and Implementation Strategies at Investment Research. He previously worked at JP Morgan, Louis Capital Markets and BNP Paribas, can received an MS and BA from Bocconi University in Milan with a specialization in Econometrics and Quantitative Modeling, attending exchange programs from ESSEC (France) and University of California, Berkeley - Haas School of Business (USA). He taught several seminars on hedge funds and quantitative strategies and was Chairman of a research workshop based on hedge funds in collaboration with EAP-ESCP University of Paris and Ca' Foscari University of Venice.
Mr.Andreani's latest project is the Klimen II System, a systematic trading program developed for futures markets. The approach is based on the analysis of price patterns to identify efficient entry and exit points; Data Mining and Artificial Intelligence methods are used to select the most robust trading strategy for intraday trading, depending on market conditions. The program has been performed on liquid futures based on Bonds, Equity Indices and Forex. The system is implemented on a proprietary platform which runs the strategies systematically.
Opalesque Futures Intelligence (Mark Melin): Short-term time duration quant traders are not entirely commonplace in managed futures. Usually the strategy is associated with high frequency proprietary trading. Certain allocators are looking to fill an ultra-short term time slot, so this is fascinating. Tell me more about yourself and how your program developed.
Alexandre Andreani: For the last ten years I have been designing and managing my proprietary quantitative trading strategies in low and high frequency models. I have led and managed R&D projects ranking scale up to ten man years and several thousand Euro budgets to develop algorithmic quantitative investment programs and strategies that have been implemented successfully in single manager hedge funds.
Our program holds positions for a few hours in the most liquid futures contracts in US and European markets including equity, bond, Forex, futures. Efficient entry points are identified by the analysis of price based indicators constructed from behavioural principles and whose significance is assessed by extensive data-mining investigations. The effects are accounted by the analysis of out of sample consistency of trading rules.
Our trading program runs on a propriety platform fully developed in-house, allowing us to hold full control of every detail of the investment process. The platform handles the maintenance of the time series database, the development and evaluation of the pattern recognition modules, real time trading, and risk management.
We currently trade 6 instruments and our objective is to get a basket of 15 instruments in our portfolio by October.
OFI: Are your servers co-located in the exchange facility or are the servers located in Europe?
AA: We just finished our testing-phase to verify the correct functioning of the platform. We are using the cloud computing to mitigate the risks of market connection failure and improve the latency of each trade. The co-location is necessary for a systematic managed futures, we will have two co-locations: one in Europe and one in US.
OFI: When you said short-term trades, you mention that about two hours as an average hold period is that correct?
AA: It could be more, that depends on the pattern identified by our platform. It could be three hours, it could be five hours or even a fixed time of the day, but at the end of the day, we close all our positions. We are completely flat at the end of the day, for sure. This is how we designed the program. I believe intraday/short-term approach is an edge for investors. For example, if we take the Mini-S&P we close our positions few minutes before market close. The model will enter into positions during the day and once it identifies a signal, it will decide to enter into position; simultaneously it will use an outright stop order to protect the position. Plus, our model includes a volatility filter to block positions opening in conditions of high and low directionality of the market. Position limits are enforced so that the maximum exposure is fixed and the total number of trades per day is fixed.
OFI: Vol filter, interesting. Tell me about your risk management. Can you explain to me how you enter and exit trades?
AA: We have two different levels of risk management protocols: real time risk management and end of the day risk management.
Our real time operational risk relies mainly on outright stop orders submitted for each open position. Furthermore, we have position limits coded in the logic of each basket strategy following the portfolio sizing. When the maximum position size is reached, no more trades will be generated.
"...we have a tick filter which detects high volatility alerts when the difference between two consecutive ticks crosses a predefined threshold."
In addition to broker filter, we have a tick filter which detects high volatility alerts when the difference between two consecutive ticks crosses a predefined threshold. We also have the market access connection failure which generates specific alerts. This situation, which might require the intervention of our trader if the platform failed to re-establish the connection automatically. This is part of the real time risk management protocols that are completely coded and automated in the system. Our end of the day risk management protocol is designed to control risk from the analysis of the daily performance of the program.
Such analysis is carried out at the end of the each day with the validation of cumulative daily performance drawdown and its comparison with certain risk parameters such as, for example, the VaR levels for its strategy and for the whole portfolio.
This is the way how our risk management process is functioning.
OFI: Typically, investors want to understand where the risk is in an investment. There is risk in all investments. Let's talk about the macro situations under which this strategy performs well and then after that give me the scenarios under which the strategy might experience difficulty.
AA: When you are running an automated short term trading strategy it could be the worst case scenario and the relative length to recovery for the model. For example, when considering the worst case scenario of our model, a critical aspect is, I believe, is assessed when market volatility is excessively high, and at this point the number of operations should drop and the strategy should be stopped. In the inverse, when volatility is extremely low, you should also monitor the type of switching regimes and the behaviour of the model in this kind of scenario. This is a factor to consider. So, I believe the most important aspect to monitor is the recovery time, which would determine the total duration of the drawdown.
"...when volatility is extremely low, you should also monitor the type of switching regimes and the behaviour of the model in this kind of scenario."
In that respect, I believe one interesting topic to analyze is the daily consecutive loss probability and the distribution probability. It may show the frequency of consecutive daily loss based on backtested data. We really care about the frequency of daily percentage returns on different time horizons. For example, we need to observe the frequency of consecutive daily loss since it will determine how many days and how many consecutive loss days we need to monitor to re-establish a normalcy situation. Then, of course, you have to go deeper into that analysis and scrutinize the frequency of daily returns.
Our objective here is to effectively figure out the capability of the program to perform in special market situations as, for example, a switch of volatility regimes or a sideways movement. Our job is, first to fine tune the program in order to mitigate the probability of unfavourable events to enhance stability and robustness during the next time window of observation. Second, to continuously monitor the consistency of the performance and detect signs of style drift. Another aspect to control is how many trades the model performs per day. When volatility is extremely high, the number of signals could be higher than the average and this could cause a misinterpretation of the model with respect to the market condition. I do not want to be exposed to these kinds of risks and situations.
OFI: How did you conduct correlation analysis?
AA: What we do is generally analyze empirically the daily correlation of each trades. We find out that positions taken throughout the day are quite consistent. I interpret the empirical studies as indicators of stability to check if the model is compliant and genuine versus its back-test. I do not want to be exposed to a model misinterpreting with too many different trading positions during the day. It is statistical probability trading. Our objective is to present on the market the day the model will best interpret the market. For me, these kinds of daily correlations are very important to determine signs.
OFI: Discuss the indicator from a stability standpoint a bit, describe it.
AA: It is the average number of distinct trader automation used on different time scales. The topic relies on what is the percentage of trading automation using two consecutive time scale? We list for each product the percentage of trader automata, which are maintained from a year to the following one, together with the average number of distinct patterns, which are used in a year. We have data on monthly and quarterly frequencies. It is very interesting to check how the model will react in terms of changing and switching scenarios.
OFI: You mean volatility scenarios?
AA: Yes. For example, if we are in a state of high volatility for a quarter, how the model will interpret the next quarter, the next observation.
OFI: What you are saying is the system works best in a high volatile environment where price persistence exists in one direction and then the risk is when the volatility is not there and the intraday place persistence is not there.
AA: Exactly. The point here is that the model should be adaptive and self learning. Our protocol of trader automaton construction is designed to take advantage of the information coming from new market conditions. The impact of new data on the investment process is weighted against the signature of past events, with market shocks events having a stronger effect in modifying existing pattern recognition scheme. This framework enables us to be able to improve the performance with time, as the program learns from experience.