COSMOS Hedge Funds
and Trading Expert-System TEXSOL
A White Paper
April 2005 - DOI: 10.13140/RG.2.2.15019.75044
Patrice Poyet and Guillaume Besse
COSMOS stands for Capital Oriented Systematic Management of Speculation. The key objective is to operate
speculative hedge funds based on systematic trading and automated account management by expert systems
while favouring protection and growth of capital. All words have their importance. Speculation means that the
approach does not deal with any notion of absolute or relative value of the securities or of the sectors or markets
addressed (i.e. fundamental analysis) and that tight risk control mechanisms are enforced to ensure that the
accounts be managed according to the rules enacted, the policies described to and agreed by the stakeholders.
Methods used to trade have been completely formalised and not a single decision is made by the system which
would not correspond to a clear trading rule, checked and back-tested on many markets over decades. The notion
of trading rule is very generic at that stage of this introduction and covers mechanisms enabling to identify
markets, sectors or securities’ trends, to determine long / short entry signals, to handle positions sizing and
management including stop policies, to care for account and money management within agreed boundaries or
safety limits (e.g. draw down, etc.).
The method and the expert system implemented (TEXSOL – Trading Expert Systems On Line) enable complete
and objective reporting of the “why” and the “how” of each and every decisions taken while running the
accounts. This guarantees the stakeholders in the funds that no discretionary decisions be made by operators and
traders running the funds and that all actions have a sound basis and correlated justifications. Proper trading
remains a way of applying methods, techniques and tools in order to maximise profit expectancy while
minimising risk but the future COSMOS fund stakeholder should understand that while the methods and expert-
system used ensure coherency and convergence of the results over reasonable periods of time, trading – even by
extremely consistent automated means and methods – remains over the short term an activity which draws on a
distribution of probabilities as generated by the methods used, with non deterministic results for any single trade.
Furthermore, one should understand that any single event, such as adding or removing trading capital from the
funds leads to a different succession of decisions and events given the impact on timing, added (or withdrawn)
leverage and on the selection of securities and positions taken or rejected, generating different alternatives – each
acting as parallel universes - with slightly different returns.
While over increasing periods of time and with given risk control parameters the system delivers coherent and
consistent performances with small deviations, one should understand that over short timescales (i.e. typically
less than two years) significant deviations might arise and do happen from time to time. COSMOS has been
tested and operated daily on many hedge funds trading universes, including stocks on the US market (i.e. a
selection of 1224 securities), the French market, indexes, Exchange Traded Funds (ETFs) and other financial
instruments. Hundreds of runs have enabled the authors to assess the maximum deviations that one might face
over varying timescales, shifted timescales and many other factors having an impact over the returns. Complete
benchmarking is available upon request and graphically represents the distribution of the yearly returns against
the maximum Peak-to-Through encountered over increasing sliding windows of 24, 36, 48, 60 months, the
yearly returns against increasing holding duration and the impact of the initial trading capital on the yearly
returns, this for the COSMOS US trading universe of over 1200 securities selected among six technology
The method and the system developed are definitely of the “hedge fund” type. COSMOS results should not be
mainly assessed by comparison with a benchmark (for example using an index – though this would be very
favourable to COSMOS), as the objective is to deliver positive and consistent returns, for well determined risks
levels implying appropriate protection of the principal, under all market conditions.
The approach is not market neutral, but mainly directional – positions being either “long” or “short” depending
on the system’s assessment of the prevailing trend. TEXSOL can also remain out of the markets when the trend
is not strong enough and therefore when the risk of having positions whipsawed is greater than the profits to be
expected. Bull market corrections are not addressed by any hedging mechanism like shorting ETFs or other
synthetic, future or derivative product but rather by a dynamic draw down control (D3C) approach which ensures
that while the market decelerates the system would first acknowledge the situation by a self analysis of its
internal state (e.g. analysis of the potential open trades) and eventually would lighten up portfolio positions
according to “garbage collection” mechanisms “GC”.
During Bull market accelerations, the system uses full leverage on cash to load on positions on margin, while
during “neutral” Bull market conditions the system limits its actions to basic positions and portfolio
management. The importance of this “swing trader like” self-adjusting behaviour depends on the risk parameters
specified for the D3C during account activation. As a general rule, increasing control and decreasing risks also
tend to decrease returns, though the relationship is neither naïve nor linear.
Bear markets are addressed through “shorting episodes” which are handled through “opportunity windows”, and
do not require the D3C facility to be activated but rather a close control on portfolio stops posted against the
equity curve of the fund. The trading universes must be built on liquidity assumptions and requirements to
ensure that proper liquidation of positions be realistic under most market conditions, while consequences of
more abnormal market situations can be measured through VaR, market stress and market shock scenarios.
Operating the method and system
On a daily basis, the system updates the database of its various trading universes and computes all technical data
required for the securities, their GICS sectors (i.e. Global Industry Classification Standard from S&P and MSCI)
and reference indexes and stores the results in a data storage referred to as “historical database”. For each
account / fund – and given the associated trading universe, its strategy and various risk control and portfolio
parameters – the system generates the potential open trades according to the prevailing market trend it has
identified, all information being stored in the “account database”.
These open trades are then evaluated and ranked according to various technical indicators (e.g. RSC) and a mark
reflecting the ratio “profit expectancy” on “risk” - expectancy being derived from various information including
statistics compiled by GICS sectors on past trades in similar sectors for various time windows (with exponential
moderation over time) and risk being assessed mainly based on initial stops (adjusted against volatility on shorts
as measured by the ATR).
Given the market trend and several market internals, the system will adopt a behaviour such as add on positions
if possible, just handle current positions, or lighten up according to various GC mechanisms. Adding new
positions requires proper verifications on the portfolio level, to ensure that various draw down(s) levels will not
be exceeded, that leverage is not bypassed potentially leading to undesirable margin calls and therefore
jeopardizing current positions, etc. When all verifications have been made, stops adjusted, positions checked,
new positions are eventually taken (buy, sell, sell short, buy to cover), and a report delivered so that
corresponding orders be updated on the market.
The portfolio as a whole is also monitored and the dynamics of the equity curve checked to post if necessary
stops against the valorisation of the entire account to ensure proper global exit when the trend ends to be a friend
and market reverses, whatever the direction up (while net short) or down (while net long). Uploading from the
TEXSOL account database in Excel enables a daily control and reporting of all account parameters and positions
for each account / fund under management.
Putting it all together
The method designed and the system implemented enable to create custom trading universes satisfying particular
objectives (e.g. sector oriented, ETF based, or any other criterion) and to handle specific accounts / funds
according to specified risk profiles. The authors have been using their money and own accounts as test-accounts
showing similar logs and returns as to what is observed during extensive back testing over more than a decade on
the US and French markets. Deployment of the method and technology is very flexible and can be arranged at
standard industry rates and conditions, i.e. 2% fees for funds under management, and 20% incentive fees on
returns past high watermarks. Authors are open to any deals with investors, institutions and broker dealers.
Yearly returns can be as high as 40% on a long term basis (typically over the 1995-2004 period) for technology
laden trading universes leading to remarkable performances for well mastered risk policies, though 20-25% are
more reasonable expectations.
Patrice Poyet and Guillaume Besse, 2005. Systèmes Experts de Trading en Ligne - Trading Expert Systems On
Line TexSOL. DOI: 10.13140/2.1.1212.4809 - March 2005 - 194 pp.