A derivatives trading recommendation system: The mid‐curve calendar spread case

• Date: 01 April 2019
• DOI: http://doi.org/10.1002/isaf.1445
• Published date: 01 April 2019
• Author: Adriano S. Koshiyama,Nikan Firoozye,Philip Treleaven

Received: 18 February 2018 Revised: 12 November 2018 Acc epted:11 March 2019

DOI: 10.1002/isaf.1445

RESEARCH ARTICLE

A derivatives trading recommendation system: The mid-curve

calendar spread case

Adriano S. Koshiyama Nikan Firoozye Philip Treleaven

Department of Computer Science , University

CollegeLondon, Gower Street,London, WC1E

6BT, United Kingdo m

Correspo ndenc e

AdrianoS. Koshiyam,Department of Computer

Science, Un iversity College London Go wer

Street, London W C1E 6BT, United Kingdom .

Email: as.koshiyama@gmail.c om

Funding information

National Research Co uncil of Brazil

Summary

Derivative traders are usually required to scanthrough hundreds, even thousands of

possible trade s on a daily basis . Up to no w, not a sing le solution is available to aid in

theirjob. Hence, this workis aimed to develop a trading recommendationsystem, and

to apply this system to the so-called Mid-Curve Calendar Spread (MCCS) trade. To

suggest that such approach is feasible, we used a list of 35 different types of M CCSs;

a total of 11 predictiv e and 4 bench mark models. Our results su ggest that linear

regression with l1-regularisation (Lasso) compared favourably to other approaches

from a predictive and in terpretability point of views.

KEYWORDS

derivatives, mach ine learning, trading recommendation sys tem

1INTRODUCTION

Derivativetraders areusually required toscan through hundreds, even

thousands of possible trades on a daily basis. A concrete case is the

so-called Mid-Curve Calendar Spread (MCCS), a derivatives pac kage

that involves selling anoption on a forward-starting swap and buying

an option on a spot-starting swap with longerexpiration (Corb, 2012;

Natenberg, 2014). In such a package, traders look for the historical

carry and the breakeven width levels, metrics that can be easily

inferre d from the term inal or aged pay off profi le of the MCCS , shown

in several heatmaps made by the research team. Afte rthat, they rank

the most prominent ones to offer a client or to proceed in some

proprietarytrading. In general, the straightforwardness and swiftness

that the decisions are mad e is the main upside of this framework.

However, o ne might notice that the m ain downside s of such

approach are: (i) substantial info rmation on the und erlying like sen-

sitivities, implie d volatility, etc . are usually n ot taken into acc ount;

(ii) using the previous ex ample, high h istorical values fo r carry and

breake ven widt hs are rather mo re necess ary than suf ficien t condit ions

for a profitable MCCS trade, being such argument extensible to other

trades as we ll; (iii) a trader can quic kly judge if an individual pos ition is

worthwhile to inv est, but may take some time to find it; and (iv) afte r

a given period, traders tend to only lookat a small subset of possible

trades (small area on the heatmap), rather than the all available se lec-

tion. Hence, asystematic approachwhere more informationat hand is

crossed and agg regated to find good trading pic ks can be highly useful

and und oubtedly i ncrease the tra der's produc tivity.

Therefore, the o bjective of this work is to dev elop a trading recom-

mendation system that can aid derivatives traders in their day-to-d ay

routine. Being more specific, our solution is based on the following

pipeline: (i) on a certain trade date, we compute metrics and sensi-

tivities related to MCCS; (ii) these metrics are feed in a model that

can predict its expected return for a given holding period; and after

repeating (i) and (ii) f or all trades we (iii) rank the trades using s ome

dominance criteria. Ou r final solution is a model- based heatmap with

the attractiveness sc ores for each MCCS trade, which ca n be offered

to the traders and salespeople on a daily basis.

In order to suggest that such approach is methodologically and

computatio nally feasible, we started u sing a list of 35 different ty pes

of MCCS regarding expiration (3-60 months), forward (3-60 months)

and swap tenure (1-8 years). For each MCCS we used 10 years of

historical data, ranging weekly from Sep/06 to Sep/16. Usually, on

each Wednesday of a week we computed the (a) carry and breakeven

for an MCCS, (b) the package s ensitivities (gamma, vega, th eta, etc.),

(c) its present value for an at the money forward strike, and (d) its

present value after h olding this package fo r a given period – allow ing

us to co mpute t he hol ding h- peri ods- ahea d return . We carri ed out

an exploratory analysis, showing the overall performance of such

trades, highlighting the major factors that drove their performance

over time.

Intell Sys A cc Fin Mgmt. 2019;26:83– 105. wileyonlinelibrary.com/journal/isaf © 2019 John Wiley & Sons, Ltd. 83

84 KOSHIYAMA ET AL

After this exploratory step , we started the mod elling stage by:(i)

fitting a predictive m odel using as the inp ut those previou s metrics,

and the subsequent annualized returns as the output (in this case,

holding 1- year-ahead the trade); (ii) making predictio ns and comp aring

different modelling methodologies by a set of performance metrics

and benchmarks. Being more precise, we used a total of 11 predictive

models, ranging from simple linearregression to supp ortv ector regres-

sion and multi-layer perceptron. We also implemented 4 benchmark

models, being two intimately linked with breakeven width and carry

level commonly used by the traders to decide which MCCS look more

attractive to a selling pitch. A fter establishing a backte sting engine

and se tting pe rform ance me trics, ou r resul ts sug gest tha t in g eneral

LinearRegression with L1 Regularization (LassoRegression) compared

favourably to other approac hes from a predic tive and interpretability

perspective.

We can summarise below a few of the contributionsand results of

our paper:

•Build a trading recommenda tion system for Euro denominate d

Mid-Curve Calendar Spread Trades (MCCS), which to the best

our knowledg e, our work is the first attempt to do that.

•A methodology that can be expandedto US dollartrades as well

as that can be replic ated to similar derivative trades but in the

equities space (calendar-spreads, butterflies,etc.).

•Ourf inalsolution, a Lasso Regression, learned a type of volatility

buying/selling strategy withoutbeing programmed to do so; its

returns distribution ac ross all MCCS tended to be right-skewed ;

and itmatched traders view on selectinggo od trades,but adding

some dynamic view on it.

In this se nse, n ext se ction p rese nts a litera ture rev iew on e xisti ng

appro ache s to retu rn/ price p redic tion /e stimat ion in d iffe rent are as

and instruments, as w ell as a brief description o f MCCS trades. Th e

third section disp lays the dataset that of MCCS trades, show ing how

the information is computed and g athered, which variables are the

input and outputs, and the main assumptions that are embedded in it.

Then, we move to modellingstrategy, highlighting themain models that

are going to be used as candidates for the recommendation system,

how they are tested and have their performance assessed. Finally,

we exhibit the results and discussions, starting with an exploratory

analysis of the dataset, and then m oving towards the performance

analysisof each mo delthrough different metrics and perspectives.We

close this work with some concluding remarks and future directions

for research.

2BACKGROUND

2.1 Related works

2.1.1 Cash instrumentsstrategies

Literature provides a growing body of evidence that price changes

can be predicted, that is, in particular circumstances and periods

securities violate the Efficient Market Hypothesis (Campbell et al.,

1997; Malkiel, 2003). This hypothesis statesthat price changes must

be unf oreca stabl e if they a re prop erly anti cipa ted, th at is if the y

fully incorporate the expe ctations and information of all participants.

Therefore, if returns can be predicted based on an information set

(e.g., historical prices, economic indicators, news, etc.), one can use

this information to trade and generate profits that are beyond w hatis

expected given the risk level thatthe market participant is assuming.

In this sense, research ers have emplo yed diffe rent modelling

approaches and information sets to predict price changes across a

range of assets. W hen we scan th e literature for cash instruments

(equities, bond s, foreign exc hange, etc .) focuse d only in using past

returns a s the mai n sourc e for pred iction, we c an find works th at tap

intoBayesian forecasting (Zhou et al., 2014), NonparametricPredictive

Inference (Baker et al., 2017), Forecasting Combination (Elliott et al.,

2013), Generalized Exponential Weighted Moving Average (Nakano

et al., 2017), Support Vector Machines (SVM) (Karathanasopoulos

et al., 2016), Shallow and Deep Neural Networks (NN) architectures

(Gerleinet al., 2016; Chong et al., 2017; Zhou et al., 2016; Deng et al.,

2017), Random Forest and Gradient Boosting Trees (Krauss et al.,

2017), and so forth. The list of proposed methodologies continue to

grow (Reveiz-Herault, 2016; Resta, 2016; Galeshchuk & Mukherjee,

2017), in which equitiesor indices appears as the dominant asset class

to apply these algo rithms. Collectiv ely, they provide e vidence that

some forecas tability of returns can be ac hieved by puttin g in place

complex models with a suitable training scheme.

Nonetheless , the main criticism of s uch approach es lies on the

limited information set thatthey are tapping into: past returns. Even

though past retu rns is a valuable source o f data for forecasting , backed

by many re ference s in the previou s paragraph, for so me authors th ey

are treated as another source of information, which in so me cases

plays a minor role for prediction. Approaches that use mainstream

and novel news sources (Reuters, The Wall Street Journal, Twitter,

etc.) have provided evidence that adding such features can improve

return prediction and aid in the de sign of a trading strategy ac ross

several asset classes (Feuerriegel & Prendinger, 2016; Chen et al.,

2014; Andersen et al., 2007; Chen & Gau, 2010; El Ouadghiri et al.,

2016). Taking an even larger information set, (Weng et al., 2017)

searches for disp arate data sources (Google, Wikipe dia, and so on) to

increase the knowledge base that an algo rithmcan tap into for making

predictions. Their inference engine used three modelling techniques

(decision trees, NN and SVM) and compared favourably to other

report ed results in th e literature.

2.1.2 Derivatives instruments s trategies

Contrasting with the emphasis that researchers in cash instruments put

on return predictability, w hen we devote our attention to resea rch in

derivatives instrumen ts (options, swaps, s waptions, etc.) it is clear that

most of the eff ort isc oncentrated on pricing th ese contracts. Similarly,

researchers have adopted different app roaches and information sets

to desc ribe th e price m ove ments o f thes e con tracts p rope rly. Th e

traditional framework is via sto chastic calculu s, in which pricing is m ade

under some market assumptions (frictionless market, no arbitrage,

risk-neutralinvestor, etc.) as well as an assumed asset p ricedynamics

over the period (e.g., Geometric Brownian Motion) (Shreve, 2004;

Glasserman, 2013). By discovering the fair price of a contract, trading

strategies can be establis hed to tap into any potentia l mispricing

(Ehrman, 2006; Natenberg,2014).