Futures Trend Strategy Model Based on Recurrent Neural Network

Transcription

1 Applied Economics and Finance Vol. 5, No. 4; July 2018 ISSN E-ISSN Published by Redfame Publishing URL: hp://aef.redfame.com Fuures rend Sraegy Model Based on Recurren Neural Nework Ru Zhang 1, Chenyu Huang 2 & Shaozhen Chen 1 1 Finance Deparmen of Inernaional Bussiness School, Jinan Universiy, Zhuhai, Guangdong Province, China 2 Financial Managemen Deparmen of Inernaional Bussiness School, Jinan Universiy, Zhuhai, Guangdong Province, China Correspondence: Shaozhen Chen, Finance Deparmen of Inernaional Business School, Jinan Universiy, Qianshan Road 206#, Zhuhai Ciy, Guangdong Province, Pos No , China. Received: May 26, 2018 Acceped: June 16, 2018 Available online: June 19, 2018 doi: /aef.v5i URL: hps://doi.org/ /aef.v5i Absrac In recen years, quaniaive invesmen has been widely used in he global fuures marke, and is seady invesmen performance has also been recognized by domesic fuures invesors. his paper akes he CSI-300 sock index fuures as he research objec and consrucs a fuures rend sraegy model based on recurren neural nework. Furhermore, his paper back ess he sraegy a differen periods, differen ransacion coss and differen parameers. he resuls show ha he sraegy model has srong profiabiliy and robusness. Keywords: fuures rend sraegy, recurren neural nework, quaniaive rading 1. Inroducion Quaniaive invesmen is defined as he combinaion of invesmen ideas or heories of invesors wih advanced compuer echnology and various disciplines, which were ransformed ino mahemaical models. hen, specific invesmen decision analysis and implemenaion will be done by compuer. In quaniaive invesmen, rend sraegy is one of he mos common models. Shik e al. (2007) forecased he role of indicaors in he foreign exchange marke hrough he saisics sudies and analysis of wo rend indicaors: RSI and MA. Elsevier e al (2009) proposed ha rading mehods based on hisorical price flucuaions and radiional echnology analysis are able o bring more benefis. Yumin Zhou (2015) upgraded he reurnee ransacion mehod by upgrading filers and delising and opimizing parameers. A he same ime, he obained a significan increase in back esing indicaors in he esing of gold fuures. Chen e al. (2016) used he LASSO mehod o selec he smalles echnical index as he judgmen funcion, and used he geneic nework plan o find he appropriae sale poins. hen, he consruced he sraegy so as o obain more benefis. Li e al. (2017) sudied he performance of buy-hold sraegy and mobile average sraegy in single variey fuures and grouping combinaions, and found ha he income of mobile average echnology is significanly higher han ha of buy-hold sraegy which had appears in China fuures marke. However, he radiional rend sraegy sill has weaknesses such as shor and weak rend and non robus parameers, which can easily lead o ransacion errors ha would be caused by repeaed ransacions and model failure. Chi e al. (1991) found ou ha recurren neural nework would remember he previous informaion and apply i o he curren calculaion; herefore i has a good predicion effec on he price rend of sock index fuures. eni e al. (2017) concluded arificial neural neworks, including RNN, are efficien in he predicion of financial ime series, and have carried ou a comparaive experimen on he predicion of foreign exchange prices. o sum up, his paper akes he CSI-300 sock index fuures as he research objec. Firs, we classify he raining daa hrough he rend sraegy profi index, sudy he parameer RNN and hen consruced he fuures rend sraegy based on recurren neural nework, which is esed by he minue marke daa for is profiabiliy. Secondly, under differen ransacion coss and policy parameers, he sraegy is compared o es is robusness. hroughou he enire research, his paper hopes o build a quaniaive invesmen sraegy in deep learning o provide furher improvemen for he applicaion of arificial inelligence heory in he field of invesmen. 95

2 2. heoreical Model 2.1 Recurren Neural Nework Model Basic Principles of Recurren Neural Nework he recurren neural nework algorihm is very good a handling daa ha depend on iming. As shown in Figure 1, in he radiional forward neural nework, he ransmission of informaion from inpu layer o hidden layer, and o oupu layer, are fully conneced. However, he nodes a differen imes are unconneced, and he nework canno deal wih he iming problem direcly. However, in recurren neural nework, here are connecions beween he nodes of he hidden layers a differen ime. he curren oupu of a sequence is depended on he previous sae, which means ha he nework will memorize he preceding informaion and apply i o he calculaion of he curren oupu. Figure 1. Recurren neural nework schemaic RNN can be regarded as a neural nework ha shares weighs on ime dimension. he oupu of he RNN model is as follows: y σ ( W h b ) (1) 0 y 0 he oupu of he model depends on he hidden sae h a ha ime and he hidden sae h depends on he inpu x of he ime and he hidden sae h 1 a he previous momen, ha is: h σ ( W x W h b ) (2) h x h 1 n herefore, he implici sae is ime dependen, so he oupu of he RNN model is relaed o he inpu informaion of he previous ime. Figure 2. RNN Parameric schemaic diagram he relaionship beween he RNN model and he ordinary ime series model can be analyzed in he following way, which considers he single variable siuaion: x x, W h β, W x α. When oupu y h, assuming ha σ h is an ideniy ransformaion, he RNN model can be ransformed ino he following forms: y αx βy b (3) 1 In essence, he model RNN can be regarded as a firs-order auoregressive model wih exogenous variables, so i is a complex nonlinear ime series model. 96

3 2.1.2 Parameer Opimizaion of RNN Model Gradien of he recurren neural nework can be obained hrough Back Propagaion hrough ime (BP) Algorihms Figure 3. RNN parameer learning diagram In he RNN model, since he implici sae h of he RNN is affeced by he implici sae h 1 a ha ime, he gradien of he RNN is relaed o he sequence. he BP is based on he sequence-dependen assumpion of he RNN. he loss funcion is se a each sequence poin, and he final loss funcion L is obained as he gradien δ of he implici sae a each sequence poin: L L (4) 1 δ (5) h According o he final loss funcion, he gradien calculaion formula of each parameer of RNN can be furher deduced, which is: 2.2 rend Sraegy Model σ y y c c σ c σ ( y y )( h ) V V σ V h W h W 2 diag(1 h ) δ ( h 1) 1 1 h b h b 2 diag(1 h ) δ 1 1 h U h U diag(1 h ) δ( x ) his sudy used he following rend sraegy profiabiliy indicaors as he original signal generaor for rend judgmen, expressed as: R Close Open (7) High Low R represens he proporion of he eniy par of he Candle line on he same day. In general, when R is relaively large, he rend sraegy is easier o ge profis; oherwise, he rend sraegy is more difficul o be profiable. (6) 97

4 2.3 Fuures rend Sraegy Model Based on Recurren Neural Nework As can be seen in Figure 4, marke daa of he parameer lengh is aken afer he opening dae of each rading day, and he RNN model is used o predic wheher he rend sraegy of he day can be profiable or no. If i is profiable, he rends in he early morning is followed for rend racking purposes; if he model deermines ha i canno make profis, no rading will be open on ha day. Figure 4. Schemaic diagram of sraegy Before rading, he RNN predicion model raining is required. his sudy adoped he "ime series inpu - single label oupu" model srucure. ha is o say ha he inpu daa is a mulivariae ime series, and he oupu daa is 0, 1 ag daa. In model raining, early-morning quoes for he daa in he sample need o be annoaed, and he marke is marked as wo differen caegories of suiable for rend ransacions and no suiable for rend ransacions which means ha he rend sraegy indicaor is calculaed o deermine wheher he rend sraegy of he day is profiable: when R 0.5, i is profiable, and recorded as "1" (for rend rading); when R 0.5, loss, and recorded as "0" (no suiable for rend rading). he labeled sample daa is used o rain machine learning models. Afer he open rade, he posiion is held unil he close of he day, and if a sop is riggered, he posiion will be closed immediaely afer he sop loss is riggered. 3. Empirical Analysis 3.1 Daa Selecion and Descripion (1). One-minue price of he CSI-300 Sock Index Fuures (IF) main conrac. he daa from April 16h, 2010 o December 31s, 2013 are sampled. he RNN forecasing model is rained using he marke daa for his period of ime; while i is he ou-of-sample reurn inerval from January 1s, 2014 o July 31s, (2) he characerisic daa sequence is he minue-level marke daa of he opening price, he closing price, he highes price, he lowes price, he rading volume, he main buying amoun, he main selling amoun, he change rae of he closing price, and he closing price sequence of he minue marke, order differenial, raio of main buyer sales, volume change rae, change rae of main purchase volume, change rae of main sales volume. Feaure daa is sandardized. 3.2 ransacion Parameer Seings (1). Back es margin seing: 100%, ie no leverage is aken ino consideraion. (2). ransacion coss: 100%, ie no leverage is aken ino consideraion. (3). Sop loss seing: sop loss according o a fixed raio 3.3 Model Evaluaion Index his sudy selecs he annualized ineres rae, he cumulaive yield, he maximum reracemen, he winning raio, he profi-loss raio, and he single-ransacion yield as indicaors of he evaluaion sraegy. 3.4 Simulaion Back esing his sudy seleced 33-minues afer he opening o forecas rend sraegy profi index R of he full-day. he rend of he profiabiliy probabiliy P is shown in Figure 5, and he ou-of-sample forecas accuracy rae is 59.1%. 98

5 Figure 5. Profi Probabiliy Char he performance of he sraegy model across he enire sample inerval is shown in Figure 6. Since April 2010, he sraegy has been raded a oal of 707 imes, achieving a cumulaive yield of %, an annualized rae of reurn of 18.01%, a maximum reracemen of -8.63%, and a profi-o-loss raio of able 1. Sraegy Performance in 2010 Figure 6. Full sample performance Index Value ime yield 18.01% Cumulaive reurn % Maximum reracemen -8.63% ransacion number 707 odds 40.74% Profi and loss han 2.17 Yield per rade 0.17% his sraegy model is a rend rading sraegy wih a sop loss mechanism, so he winning rae is no high, which is 39.52%, bu he profi and loss of he sraegy is high a 2.27%. Sraegy of a single ransacion is an average yield of 0.17%. Compared wih oher indexes of CA shor erm sraegy, he sraegy of a single ransacion average yield is higher. herefore, he sensiiviy o ransacion coss and marke impac coss is low. Since Sepember 2015, sock index fuures rading have been grealy resriced, ransacion coss increase, marke liquidiy becomes worse and impac coss increase. herefore, he average reurn rae of a single ransacion of CA sraegy is an imporan index. If he ransacion coss raise o bilaeral five over en housand, he performances of he sraegy ouside he sample are shown in figure 7. In he case ha ransacion coss sharply raised, he performance of he sraegy is slighly lower. However, since 2014, he annual yield of he sraegy is 14.71% and a single rading average yield is 0.14%. Generally, he 99

6 performance is sill good, which means ha his ransacion sraegy is no sensiive o ransacion coss. able 2. Sraegic performance in 2014 Figure 7. sraegic ne worh a differen ransacion coss (ou of sample) Index Value Annual reurn 18.47% Cumulaive rae of reurn 80.72% Maximum wihdrawal -8.63% Number of ransacions 372 Winning rae 39.52% Profi and loss raio 2.27 Single ransacion yield 0.17% 3.5 Parameer Sensiiviy Analysis he opening ime is he main parameer of his sraegy, and 33 minue is he opimal opening ime parameer obained from he sraegy in his repor. In order o analyze he sensiiviy of parameers, differen were seleced for back measuremen. he specific sraegy is he same as above; he daily rend sraegy profi probabiliy P is prediced based on he marke daa afer he opening minues, and hen he daily 120 average daily of he probabiliy P a ime is calculaed. If P MA ( P) 120, hen rend rading is carried ou. According o he main conrac of IF afer he opening minues, he rend direcion is esablished from he direcion of flucuaion; Oherwise, do no open a posiion on he day rading. Under differen parameer seings, he performance of he sraegy ouside he sample is shown in figure 8. Any ime sraegy ha is performed during 22 minues o 38 minues afer opening o rade, i shows good performance; herefore i explains ha his sraegy has good sabiliy. Figure 8. Performance of sraegy a differen warehouse opening imes (ou of sample) 100

7 able 3. Performance of sraegy a differen warehouse opening imes (ou of sample) Parameer (min) Annualized rae of reurn Maximum reracemen Parameer (min) Annualized rae of reurn Maximum reracemen % % % -8.02% % % % -8.06% % % % % % % % % % -9.71% % % % -9.47% % % % -8.89% % % % % % % % -8.87% % % % % % % % % % % % % % % % -8.15% % % % -8.63% % % % -8.71% % % % -8.90% 4. Conclusion his sudy adoped he sock index fuures marke in early rading. A predicive model is esablished by using a recurren neural nework in order o evaluae he rend and shock condiion of he marke and forecas he probabiliy of daily index fuures marke rend sraegy profi. Based on his probabiliy, he suiabiliy of he day for rend rading is deermined and wheher o open a posiion is also decided. Empirical shows ha hrough he rading sraegy, smaller rading signal wih smaller profi opporuniies will be filered while he sraegy will only perform rend rading if probabiliy of profi is high. In his way, his sraegy significanly improved he profiabiliy of he policy. In addiion, he empirical resuls show ha he sraegy has good parameer sabiliy hrough differen ransacion cos seings or differen parameers of he policy rees. Reference Chong,. L. (2007). A comparison of MA and RSI reurns wih exchange rae inervenion. Applied Economics Leers, 14(5), hps://doi.org/ / Chi, S. C., Chen, H. P., & Cheng, C. H. (1999). A forecasing approach for sock index fuure using grey heory and neural neworks[c]// Inernaional Join Conference on Neural Neworks. IEEE, 1999: , /IJCNN Li, B., Zhang, D., & Zhou, Y. (2017). here rends in Chinese Commodiy Fuures Marke. Securiies Marke Herald, (1), Paoloeni. (1996) Forecasing foreign exchange raes using recurren neural neworks. Applied Arificial Inelligence, 10(6), Venna, J., Kaski, S., & Pelonen, J. (2003). Visualizaions for Assessing Convergence and Mixing of MCMC[M]// Machine Learning: ECML Springer Berlin Heidelberg, hps://doi.org/ / _39 Yu, M. Z. (2015). Developmen and improvemen of fuures program rading sysem based on rading pioneer and urle rading rules, China Marke, 2015(29), Yan, C., & Xuan, C. W. (2015). A Sudy on High-Frequency rading Sraegy Based on Variable Selecion and Geneic Nework. Chinese Journal of Managemen Science, 23(10), hps://doi.org/ /j.cnki.issn x Copyrighs Copyrigh for his aricle is reained by he auhor(s), wih firs publicaion righs graned o he journal. his is an open-access aricle disribued under he erms and condiions of he Creaive Commons Aribuion license which permis unresriced use, disribuion, and reproducion in any medium, provided he original work is properly cied. 101