Oil Price Trackers Inspired by Immune Memory

Oil Price T rackers Inspir ed by Immune M e mory W illiam W ilson * , Phil Birkin * , and Uwe Aickel in * * School of Computer S ci ence, Univ ersity of Not ting ha m, UK wow, pab, uxa@ cs. nott.a c.uk Ab st r act W e ou tline initial c oncepts for an i mmune inspire d algorithm to evaluate an d predict oil p rice time se- ries data. The pr oposed solution evo lves a short ter m pool o f tr ackers dynami cally , with eac h member attempting to map trends a nd anticipate future price mov e ments. Successful trackers fe ed into a long term mem ory pool that can generalise acr oss repeating trend patterns. Th e res ulting se quence o f track- ers, ordered i n time, can be used as a forecasting tool. Exami nation of the pool of ev olving trackers also provides v a luable insight into the properties o f the crude oil market. 1 Intr oduction The in vestigation of ti me serie s data to predict future information is a well studie d ar ea of researc h. This paper proposes a n imm une i nspired sol ution to thi s problem. I nspiration f or memor y deve l opment was taken fr om the biol ogical theory proposed b y Dr Eric Bell. The t heor y i ndicates the existence of two sep- aratel y identif iable memor y populations (Bell, 2 005), one long term and the o ther short ter m. Their d if- fering characterist ics make them ideal i n rec ognising long and short term trends prev alent in time series data. These tren ds can t hen be seq uenced f or u se in forecasting an d prediction. 2 De velopment o f long and short term memory The flexible learning approach of f ered b y t he im- mune s y st em is attrac tiv e as an inspiration for pr ob- lem solvin g. Ho wever without a n a dequate memor y mechanism the kno wled ge gained fr om the learni ng process would be lost. Memo ry ther efore r epresents a key contributing f act or in the succe ss o f t he im- mune s y ste m. A dif ficult y a rises i n extra cting im- mune memor y propertie s howev er, because v er y lit- tle is still known ab out all the biological mecha nisms underpinnin g me mor y development (W il son a nd Gar- rett, 2004). Theo rie s such as antigen pe rsistence and long liv ed memor y cells (Perelson and W eisbuch, 1997), idiot ypic networks, a nd home ostatic t urnov er of memory ce lls (Y a tes and Callar d, 2001) hav e all at- tempted to explain the dev elo pment a nd maintenanc e of immune memory but all have been conteste d. The attraction of the im mune memor y theory pr o- posed by Dr Eric Bell is that it p r ovides a si mple, clear and l ogical explanati on of memory cell de velop- ment (Bell , 2005). This theory h ighli ghts the evolu- tion of t wo separate me mory p ools. T he fir st is a short term memory pool contai ning sho rt liv e d, highl y pr o- liferativ e, a ctiv ated cells that hav e experience d an antigen. The purpose of this pool is t o driv e the affin- ity m atura tion process to cop e with the huge diver - sity of potentia l anti gen mut ati o n s. The second p ool consists of th ose short term memory cells that hav e ev olved to homeostat ically tur nover to sustai n k nowl- edge o f a n a ntigen experienc e over t he l ong term. This long ter m pool ident ifies and maintai ns kno w l- edge of more generalised a ntigen tre nds. 3 Analysis of oil price tr ends The price of W TI crude oil (a world marker price f or oil price m ov ement s) was sel ected as t he time series for i n vestigation. Th is data s et was chose n be cause there is c onsiderable economi c, financia l and gov ern- ment interest in inv e stigating oil price foreca sting due to its influe nce on so many other mar k et sect ors. In addition, oil price s h ave hist o r icall y exhibited a num- ber of sh ort and long ter m trend patter ns which could map to our long and sh ort term memor y c oncepts, providing an ideal case study for this analysis. 4 An immune inspire d f o re cast- ing solution The pr oposed solution comprises a population of ”trackers” that corre spond to B cells from the im - mune s ystem. The t rackers at tempt t o iden tify and record trends in the oil price data. Pr ice data, a s mea- sured b y t he change in price fr om one ti me peri od to t he ne xt, is encaps ulated within a n art ificial anti- gen object an d presente d to t he populati on of trac kers. The ant igens are constr ucted to sh o w current and his- torical price m ov ement s ov er a par ticular period. In order to recognise p rice trends over t ime, each tracker is allocated a random l ength ”rev iew period”. This allows the tracker p opulation to i dentif y a va r iet y of potential price mov eme nts over a range of ti me inter- va ls. Follo wi ng the tra ditional clonal selection approach (de Castr o an d V on Zuben, 2 002), t rackers a ttempt to bind to an tigens, a nd undergo prolifera tion if suc- cessful. The resulti ng clones mutate in r elation to t he strength of the bind, with mutation ta king one o f t hree forms. One subset of clones has a random price v alue within their revie w periods m utated from its original va lue. A second subset has t heir revie w period ex- tended b y th e a ddition of a randoml y generated price mov eme nt t o ant icipate f uture potentia l pr ice move- ments. A third subset of clo nes has a ran dom pric e va lue removed from their revie w period to a ll o w them to at tempt a better fi t to previousl y experience d ant i- gens. The degree proliferati on is proportiona l t o the strength of the bind and t he length of t he boun d tracker . Initiall y trac kers have rela tiv ely short revie w periods, t o ena ble them t o asse ss a wi de variety of price trends. If successful, trac kers pr oliferate a nd the rev iew pe riods le ngthen to a nticipate addi tional price mov eme nts. E xcessiv el y l ong tra cker revie w perio ds are prev e nted because trackers bec ome m o r e specifi c as they lengthen and are there fore les s likely to bind. W ithout s uccessful b inds t hese trackers are li kely to be re mov ed via a poptosis. T his leads to the evo l ution of a dynamic population of trac k ers. The population of proliferating trac kers can be seen t o represe nt the sh ort term me mor y of experi- enced price da ta, as knowledge of an ident ified price trend is carri ed f orward thr ough the generat ions of tracker cl ones. Interr ogating th e comp osition of this memor y p ool provides valuable i nsight into the d y - namics of the oil mar k et . The process of filtering the short term memor y pool to a lon g ter m m e mor y sub set is achieved through development of t he ”‘trac ke r sequenc e”’. The tracker sequence is a list of tra ckers, ordered in time, that best r epresents the data p r esented up t o the current point in time. Dominant tracker cand i- dates, based on their degree of pr oliferation, are se- lected fr om the sh ort ter m mem ory pool a nd tra ns- ferred t o the trac ker sequence f or u se as a source of long term memor y . Ge nerali sations can be made i n the tracker seq uence for repeating pattern s of tra ck- ers t o hi ghlight recurring pri ce trends. The tracker sequence provides the f orecasting mecha nism in the system. When new pric e data bec omes av ailable the trac k er sequence is examine d to iden tify whether a prev i ousl y i dentified trend is recurring ag ai n. 5 Conclusi on Inspiration was ta k en fro m the principle s of me m- ory within the i mmune s ystem to build a s y st em that would iden tify trends within an o il price time series. This data showed evidence of short ter m price fluc- tuations a s well as exhibitin g under lying long term trends. Detailed ins pi ra tio n was taken f rom the t heor y of immu ne memor y pro posed b y Dr Er ic Bell which identifies two f orms of m emor y , sh ort ter m and l ong term. W e indica te tha t t hese could in pri nciple pro- vide a mechanis m to i dentif y and map the s hort and long term trends evident i n t he crude oi l mark et which could then be used for forecast ing. Acknow ledgements The authors would l ike t o t hank D r Eric Bell from the Univ e rsit y of Manchester for his valuable input. References E. B ell. U niv ersit y of Manch ester, pers onal commu- nication, 2005. L. N . de Castr o and F . J. V on Zuben. Learning and optimizati on using the clonal sele ction princi ple. IEEE T rans actions on Evol utionary Computatio n , 6(3):239–251 , 20 02. A. S . Pe relson and G. W eisbuch. Immunol ogy f or physicists. Rev . Modern Ph y s. , 69:1219–1267, 1997. W . Wilson a nd S. Garrett. Modelling i mmune mem- ory for pred iction and com putation. In 3r d Inte r - national Confer e nce i n Arti ficial I mmune Systems (ICARI S-2004) , pages 386–399, Catania, Sicil y , Italy , Se ptember 2004. A. Y a tes and R. Callard. Cell deat h an d the main- tenance of im munological me mory . Discre te and Continuo us Dynamical S yste ms , 1 : 43–59, 2001.