Generating Realistic Stock Market Order Streams
Junyi Li, Xitong Wang, Yaoyang Lin, Arunesh Sinha, Micheal P. Wellman
GGenerating Realistic Stock Market Order Streams
Junyi Li, Xintong Wang, Yaoyang Lin, Arunesh Sinha, Michael P. Wellman University of Pittsburgh, University of Michigan, Harvard University, Singapore Management [email protected], [email protected], [email protected], [email protected], [email protected]
Abstract
We propose an approach to generate realistic and high-fidelity stock market data based on generative adversarialnetworks (GANs). Our Stock-GAN model employs a con-ditional Wasserstein GAN to capture history dependence oforders. The generator design includes specially crafted aspectsincluding components that approximate the market’s auctionmechanism, augmenting the order history with order-bookconstructions to improve the generation task. We perform anablation study to verify the usefulness of aspects of our net-work structure. We provide a mathematical characterization ofdistribution learned by the generator. We also propose statis-tics to measure the quality of generated orders. We test ourapproach with synthetic and actual market data, compare tomany baseline generative models, and find the generated datato be close to real data.
Financial markets are among the most well-studied andclosely watched complex multiagent systems in existence.Well-functioning financial markets are critical to the oper-ation of a complex global economy, and small changes inthe efficiency or stability of such markets can have enor-mous ramifications. Accurate modeling of financial marketscan support improved design and regulation of these criti-cal institutions. There is a vast literature on financial marketmodeling, though still a large gap between the state-of-artand the ideal. Analytic approaches provide insight throughhighly stylized model forms. Agent-based models accom-modate greater dynamic complexity, and are often able toreproduce “stylized facts” of real-world markets (LeBaron2006). Currently lacking, however, is a simulation capable ofproducing market data at high fidelity and high realism. Ouraim is to develop such a model, to support a range of marketdesign and analysis problems. This work provides a first step,learning a high-fidelity generator from real stock market datastreams.Our main contribution is Stock-GAN: an approach to pro-duce realistic stock market order streams from real marketdata. We utilize a conditional Wasserstein GAN (WGAN)(Arjovsky, Chintala, and Bottou 2017; Mirza and Osindero2014) to capture the time-dependence of order streams, with
Copyright c (cid:13) both the generator and critic conditional on history of orders.The main innovation in the Stock-GAN network architecturelies in two deliberately crafted features of the generator. Thefirst is a separate neural network that is used to approximatethe double auction mechanism underlying stock exchanges.This pre-trained network is embedded in the generator en-abling it to model order processing and transaction generation.The second feature is the inclusion of order-book informationin the conditioning history of the network. The order bookcaptures the key features of market state that are not directlyapparent from order history segments.Our second contribution is a mathematical characteriza-tion of the distribution learned by the generator. We showthat our designed generator models the stock market datastream as arising from a stochastic process with finite mem-ory dependence. The stochastic process view also makesprecise the conditional distribution that the generator is learn-ing as well the joint distribution that the critic of the WGANdistinguishes by estimating the earth mover’s distance. Thestochastic process has no closed form representation, whichnecessitates the use of a neural network to learn it.Finally, we experiment with synthetic and real market data.The synthetic data is produced using a stock market sim-ulator that has been used in several agent-based financialstudies (Wellman and Wah 2017), but is far from real marketdata. The real market data was obtained from OneMarket-Data, a financial data provider. We propose five statisticsfor evaluating stock market data, such as the distribution ofprice and quantity of orders, inter-arrival times of orders, andthe best bid and best ask evolution over time. We compareagainst other baseline generative models such as recurrentconditional variational auto-encoder (VAE) and DCGAN in-stead of WGAN within Stock-GAN. We perform an ablationstudy showing the usefulness of our generator structure de-sign as elaborated above. Overall, Stock-GAN is able to bestgenerate realistic data compared to the alternatives. An ap-pendix in the full version provides all additional results andcode for our work.
WGAN is a well-known GAN variant (Goodfellow et al.2014; Arjovsky, Chintala, and Bottou 2017). Most prior workon generation of sequences using GANs has been in thedomain of text generation (Press et al. 2017; Zhang et al. a r X i v : . [ q -f i n . S T ] J un igure 1: Representation and evolution of a limit order book.2017). However, since the space of word representationsis not continuous, the semantics change with nearby wordrepresentation, and given a lack of agreement on the metricsfor measuring goodness of sentences, producing good qualitytext using GANs is still an active area of research. Stockmarket data does not suffer from this representation problembut the history dependence for stock markets can be muchlonger than for text generation. There are many advancedproposals to deal with long term dependence (Neil, Pfeiffer,and Liu 2016; Chang et al. 2017; Yu et al. 2017), however, wefind that our use of LSTMs with conditional WGAN performsquite good with little tuning of hyperparameters. Xiao et al.(2017; 2018) introduced GAN-based methods for generatingpoint processes; they generate the time for transaction eventsin stock markets. Other work aim to generate transactionprices in a stock market (Da Silva and Shi 2019; Koshiyama,Firoozye, and Treleaven 2019; Zhang et al. 2019; Wiese etal. 2019). Our problem is richer and harder as we aim togenerate the actual limit orders including time, order type,price, and quantity information.Deep neural networks and machine learning techniqueshave been used on financial data mostly for prediction oftransaction price (Hiransha et al. 2018; Bao, Yue, and Rao2017; Qian 2017; Zhang, Aggarwal, and Qi 2017) and forprediction of actual returns (Abe and Nakayama 2018). Asstated, our goal is not market prediction per se, but rather mar-ket modeling. Whereas the problems of learning to predictand generate may overlap (e.g., both aim to capture regular-ity in the domain), the evaluation criteria and end productare quite distinct. GANs have been used for generation ofcustomer buy orders in e-commerce setting (Shi et al. 2019;Kumar, Biswas, and Sanyal 2018), however, stock marketorders are much more complex with buys, sells, and cancel-lations; further we attempt to ensure realism of higher leveldynamics like the best bid and ask evolution over time. Limit order books
The stock market is a venue where eq- uities or stocks of publicly held companies are traded. Nearlyall stock markets follow the continuous double auction (CDA)mechanism (Friedman 1993). Traders submit bids, or limitorders , specifying the maximum price at which they wouldbe willing to buy a specified quantity of a stock, or the mini-mum price at which they would be willing to sell a quantity. The order book is a store that maintains the set of activeorders: those submitted but not yet transacted or canceled.CDAs are continuous in the sense that when a new ordermatches an existing (incumbent) order in the order book, themarket clears immediately and the trade is executed at theprice of the incumbent order—which is then removed fromthe order book. Orders may be submitted at any time, anda buy order matches and transacts with a sell order whentheir respective limits are mutually satisfied. For example,as shown in Figure 1, if a buy order with price $10.01 andquantity 100 arrives and the best sell offer in the order bookhas the same price and quantity, then they match exactly andtransact. As shown, the next buy order does not match anysell, and the following sell order partially matches what isthen the best buy in the order book.The limit order book maintains the current active orders inthe market (or the state of the market), which can be describedin terms of the quantity offered to buy or sell across the rangeof price levels. Each order arrival changes the market state,recorded as an update to the order book. After processingany arrived order every buy price level is higher than allsell price levels, and the best bid refers to the lowest buyprice level and the best ask refers to the highest sell pricelevel. See Figure 1 for an illustration. The order book isoften approximated by few (e.g., ten) price levels above thebest bid and ten price levels below the best ask; as theseprices are typically the ones that dictate the transactions in Hence, the CDA is often referred to as a limit-order market inthe finance literature (Abergel et al. 2016).a) x i (b) Generator (c) Critic Figure 2: Stock-GAN architecturethe market. There are various kinds of traders in a stockmarket, ranging from individual investors to large investingfirms. Thus, there is a wide variation in the nature of orderssubmitted. We aim to generate streams of orders that are closein aggregate (not per trader) to real order streams for a givenstock. We focus on generating orders and not transactions,as the CDA mechanism is deterministic and transactions canbe determined exactly given a stream of orders. In fact, wemodel the CDA as a fixed (and separately learned) neuralnetwork within the generation process. In this work, we limitourselves to limit orders as we do not have access to richerorder types such as iceberg or bracket orders.
We view the stock market orders for a given chunk of timeof day ∆ t as a collection of vector valued random variable { x i } i ∈ N indexed by the limit order sequence number in N = { , . . . , n } . { x i } corresponds to the i th limit order,but, includes more information than the limit order such asthe current best bid and best ask. The components of therandom vector x i include the time interval d i , type of ordert i , limit order price p i , limit order quantity q i , and the bestbid a i and best ask b i . The time interval d i specifies the differ-ence in time between the current order i and previous order i − (in precision of milliseconds); the range of d i is finite.The type of order can be buy, sell, cancel buy, or cancel sell(represented in two bits). The price and quantity are restrictedto lie within finite bounds. The price range is discretized inunits of US cents and the quantity range is discretized in unitsof the equity (non-negative integers).The best bid and best ask are limit orders themselves andare specified by price and quantity. We divide the time in aday into 24 equal intervals and ∆ t refers to the index of theinterval. A visual representation of x i is shown in Figure 2a. The architecture is shown in Figure 2. We use a conditionalWGAN (Mirza and Osindero 2014) with both the generatorand critic conditioned on a k length history of x i ’s and thetime interval ∆ t . We choose k = 20 . The history is con-densed to one vector using a single LSTM layer. This vectorand uniform noise of dimension 100 is fed to a fully con-nected layer followed by 4 convolution layers. The generatoroutputs the next x i and the critic outputs a real number. Notethat when training both generator and critic are fed historyfrom real data, but when the generator executes after trainingit is fed its own generated data as history. The generator alsooutputs the best bid and ask as part of x i , which is the outputcoming out of the CDA network. Recall that the best bid andask can be inferred deterministically from the current orderand the previous best bid and ask (for most orders); we usethe CDA network (with frozen weights during GAN training)to output the best bid and best ask; the CDA network serves asa differentiable approximation of the true CDA function. TheCDA network has a fully connected layer layer followed by 3convolutional layers. Its input is a limit order and the currentbest bid and best ask and the output is the next best bid andbest ask. The CDA network is trained separately using theorders and order-book data using a standard mean squarederror loss. Appendix B has the code for generator, critic, andthe CDA network that precisely describes the structure, loss,and hyperparameters.We use the standard WGAN loss with a gradient penaltyterm (Gulrajani et al. 2017). The critic is trained times ineach iteration. The notable part in constructing the trainingdata is that for each of 64 data points in a mini-batch thesequence of orders chosen (including history) is far awayfrom any other sequence in that mini-batch. This is to breakthe dependence among data points for the history dependentstock market data. We make this mathematically precise next. a) Synthetic price distribution (b) Synthetic quantity distribution (c) Synthetic intensity(d) Synthetic best bid/ask (e) Stock-GAN best bid/ask (f) no CDA network best bid/ask (g) no order book best bid/ask(h) Synthetic spectral density (i) Stock-GAN spectral density (j) no CDA spectral density (k) no order book spectral density Figure 3: A comparison of different statistics for generated and real synthetic limit orders. Additional results are in appendix.
We show how general stochastic process view of limit or-der generation provides an interpretation of the distributionthat the generator that Stock-GAN is learning. Recall that astochastic process is a collection of random variables indexedby a set of numbers. We view the stock market orders fora given chunk of time of day ∆ t as a collection of vectorvalued random variable { x i } i ∈ N indexed by the limit ordersequence number in N = { , . . . , n } , where n is the maxi-mum number of limit orders that can possibly show up in any ∆ t time interval. Following the terminology for stochasticprocesses, the above process is discrete time and discretespace (discrete time here refers to the discreteness of theindex set N ).The k length history we use implies a finite historydependence of the current output x i , that is, P ( x i | x i − , . . . , ∆ t ) = P ( x i | x i − , . . . , x i − m , ∆ t ) for some m .Such dependence is justified by the observation that recent or-ders mostly determine the transactions and transaction pricein the market as orders that have been in the market forlong either get transacted or canceled. Further, the best bid and best ask serves as an (approximate) sufficient statisticfor events beyond the history length m . While this processis not a Markov chain (MC), it forms what is known as ahigher order MC, which implies that the process given by y i = ( x i , . . . , x i − m +1 ) is a MC for any given time interval ∆ t. We assume that this chain formed by y i has a stationarydistribution (i.e., it is irreducible and positive recurrent). AMC is a stationary stochastic process if it starts with its sta-tionary distribution. After some initial mixing time, the MCdoes reach its stationary distribution, thus, we assume thatthe process is stationary by throwing away some initial datafor the day. Also, for the jumps across two time intervals ∆ t,we assume the change in stationary distribution is small andhence the mixing happens very quickly. A stationary processmeans that P ( x i , . . . , x i − m +1 | ∆ t ) has the same distribu-tion for any i . In practice we do not know m . However, weassume that our choice k satisfies k + 1 > m , and then it isstraightforward to check that y t = ( x i , . . . , x i − k ) is a MCand the claims above hold with m − replaced by k . Notethat unlike simple stochastic processes for transaction prices(or fundamental value of a stock) used in finance literature,such as the mean reverting Ornstein-Uhlenbeck process, ourstochastic process of market order has a complex randomariable per time step and cannot be described in a closedform. Hence, we use a neural network to learn this complexstochastic process.Given the above stochastic process view, we show thatthe generator aims to learn the real conditional distribution P r ( x i | x i − , . . . , x i − k , ∆ t ) . We use the subscript r to referto real distributions and the subscript g to refer to gener-ated distributions. The real data x , x , . . . is a realizationof the stochastic process. It is worth noting that even though P ( x i , . . . , x i − k | ∆ t ) has the same distribution for any i , therealized real data sequence x i , . . . x i − k is correlated with anyoverlapping sequnce x i + k (cid:48) , . . . x i − k + k (cid:48) for k ≥ k (cid:48) ≥ − k .Our training data points are sequences x i , . . . x i − k and asstated earlier we make sure that the sequences in a batch aresufficiently far apart. In light of the interpretation above, thisensures independence of data points within a batch. Critic interpretation : When fed real data, the criticcan be seen as a function c w of the realized data s i =( x i , . . . , x i − k , ∆ t ) , where w are the weights of the critic net-work. As argued earlier, samples in a batch that are chosenfrom real data that are spaced at least k apart are i.i.d. samplesof P r . Then for m samples fed to the critic, m (cid:80) mi =1 c w ( s i ) estimates E s ∼ P r ( c w ( s )) . When fed generated data (with theten price levels determined from the output order and pre-vious ten levels), by similar reasoning m (cid:80) mi =1 c w ( s i ) esti-mates E s ∼ P g ( c w ( s )) when the samples are sufficiently apart(recall that the history is always real data). Thus, the criticcomputes the Wasserstein distance between the joint distribu-tions P r ( x i , . . . , x i − k , ∆ t ) and P g ( x i , . . . , x i − k , ∆ t ) . Generator interpretation : The generator learns the condi-tional distribution P g ( x i | x i − , . . . , x i − k , ∆ t ) . Along withthe real history that is fed during training, the generatorrepresents the distribution P g ( x i , . . . , x i − k , ∆ t ) = P g ( x i | x i − , . . . , x i − k , ∆ t ) P r ( x i − , . . . , x i − k , ∆ t ) . Evaluating generative models is an inherently challengingtask, even in the well-established domain of image generation(Borji 2019). To the best of our knowledge, we are the first togenerate limit order streams in stock market that is calibratedto real data and as part of our contribution we propose to mea-sure the quality of generated data using five statistics. Thesestatistics capture various aspects of order streams observedin stock markets that are often studied in finance literature.Our five proposed statistics are1. Price: Distribution over price for the day’s limit orders, byorder type.2. Quantity: Distribution over quantity for the day’s limitorders, by order type.3. Inter-arrival time: Distribution over inter-arrival durationfor the day’s limit orders, by order type.4. Intensity evolution: Number of orders for consecutive T -second chunks of time.5. Best bid/ask evolution: Changes in the best bid and askover time as new orders arrive.For each of these statistics, we also present various quanti-tative numbers to measure the quality. Due to lack of space, Figure 4: Synthetic inter-arrival distributionin the main paper the results for price, quantity, inter-arrivaldistributions are shown only for buy orders. The results forthe other types are similar to buy type results and presentedin the appendix. We first evaluate Stock-GAN on synthetic orders generatedfrom an agent-based market simulator. Previously adoptedto study a variety of issues in financial markets (e.g., marketmaking (Wah, Wright, and Wellman 2017) and manipula-tion (Wang, Vorobeychik, and Wellman 2018)), the simulatorcaptures stylized facts of the complex financial market withspecified stochastic processes and distributions (Wellmanand Wah 2017). However, the simulator is still very basic andquite far from real market data. For example, fundamentalvaluation shocks are generated from a fixed Gaussian dis-tribution (Figure 3a) and quantity is always 1 (Figure 3b),whereas the real market data distributions can be seen to bequite non-smooth (Figures 5a- 5c). Thus, we use the outputof this basic simulator as our synthetic data (which we callas real in results below). We use about 300,000 orders gen-erated by the simulator as our synthetic data. These ordersare generated over a horizon of 1000 seconds, but the actualhorizon length is not important for synthetic data as it canbe scaled arbitrarily. The price output by the simulator isnormalized to [ − , , which is the reason for negative pricesin the synthetic data. Stock-GAN and baselines : Our first results show the per-formance of Stock-GAN (S-GAN in graphs) and comparesit to baselines, namely to a recurrent variational autoen-coder (Chung et al. 2015) (VAE) and the same network asours, except using a DCGAN (Radford, Metz, and Chintala2015) instead of WGAN. We show results for price distri- a) GOOG price distribution (b) GOOG quantity distribution (c) GOOG intensity(d) GOOG best bid/ask (e) Stock-GAN best bid/ask (f) no CDA network best bid/ask (g) no order book best bid/ask(h) GOOG spectral bid/ask (i) Stock-GAN spectral density (j) no CDA spectral density (k) no order book spectral density
Figure 5: A comparison of different statistics for generated and real GOOG limit orders. Additional results are in appendix.
Real,S-GAN Real,VAE Real,DCGANPrice 0.108 0.502 0.284Inter-arrival 0.18 0.756 0.923
Table 1: KS distances against real (synthetic)bution (Figure 3a), quantity distribution (Figure 3b), andinter-arrival distribution (Figure 4a, 4b—shown in two largergraphs for clarity). The results show that VAE and DCGANproduce distributions far from the real one. We capture thesedifferences quantitatively using the Kolmogorov-Smirnoff(KS) distance (Table 1).The KS distance is always in [0 , . We skip the KS dis-tance between quantity, which is always trivially one in thesynthetic data. The much smaller KS distance between realand Stock-GAN supports our claim of better performance ofStock-GAN compared to VAE and DCGAN.For intensity, we choose T = 100 seconds sized chunksof time and measure intensity as the number of orders ineach chunk divided by the total number of orders. Figure 3c shows that VAE completely fails to match the real (synthetic)data intensity. DCGAN has the same flat intensity throughoutand again failing to match the real data intensity completely.In contrast, Stock-GAN matches the real data intensity veryclosely. Ablation : The real and Stock-GAN generated best bid/askevolutions are in Figure 3d and 3e respectively. We performtwo ablation experiments, one by removing the CDA network(no cda) and one by removing order-book information (noob), shown in Figures 3f and 3g respectively. Differences canbe seen in best bid/ask means for no cda and no ob comparedto the real and Stock-GAN results, but the quantitative dis-tinction is in the spectral densities for these time series shownin Figures 3h–3k. The spectral density of a time series is themagnitude of each frequency component in the Fourier trans-form of the time series. The spectral density figures showsthe frequency component magnitude for every frequency onthe x-axis, which is a quantitative means of comparing twotime series. It can be seen that no cda and no ob have muchfewer higher frequency components as compared to syntheticspectral density, which can also be seen by the smoother timevariation in Figures 3f and 3g. Stock-GAN’s and the syntheticigure 6: GOOG inter-arrival distributionspectral density match more closely.
We obtained real limit-order streams from OneMarketData,who provided access to their OneTick database for selectedtime periods and stocks. The provided data streams compriseorder submissions and cancellations at millisecond granular-ity. In experiments, we evaluate the performance of Stock-GAN on a large capitalization stock, Alphabet Inc (GOOG).We also tried a small capitalization stock Patriot National(PN). After pre-processing, the PN daily order stream hasabout 20,000 orders and GOOG has about 230,000. Hence,naturally PN is not a good fit for learning using data hungryneural networks and our results for PN (shown in appendix)validate this claim.Relative to synthetic data, the real market data is very noisyincluding many orders at extreme prices far from the rangewhere transactions occur. Since our interest is primarily onbehavior that can affect market outcomes, we focus on ordersin the relevant range near the best bid and ask. Specifically,in a preprocessing step, we eliminate limit orders that neverappear within ten levels of the best bid and ask prices. Inthe experiment here, we use historical market data of GOOGduring one trading day in August 2017. Our results for GOOGfollow the same evaluation metrics as for synthetic data.
Stock-GAN and baselines : We show the performance ofStock-GAN and compare it to VAE and DCGAN variant ofour network. We show these results for price distribution (Fig-ure 5a), quantity distribution (Figure 5b), and inter-arrivaltimes (Figure 6a, 6b—shown in two larger graphs for clarity).As earlier, we capture these differences quantitatively usingthe KS distance shown in Table 2. Similar to synthetic data,the numbers reveal that Stock-GAN is able to model GOOG
Real,S-GAN Real,VAE Real,DCGANPrice 0.126 0.218 0.181Quantity 0.182 0.248 0.471Inter-arrival 0.066 0.835 0.154
Table 2: KS distances against real (GOOG)data better than the baselines. Intensity is measured in thesame way as synthetic data, except we choose T = 1000 seconds sized chunks of time due to the longer horizon ofGOOG data. Figure 5c shows much smoother intensity pro-duced by VAE and DCGAN as opposed to Stock-GAN whichis much closer to the real data intensity. Ablation : The real and Stock-GAN generated best bid/askevolution are in Figures 5d and 5e respectively. As for syn-thetic data, we perform two ablation experiments, one byremoving the CDA network (no cda) and one by removingorder-book information (no ob), shown in Figure 5f and 5grespectively. The quantitative distinction is seen in the spec-tral densities for these time series shown in Figures 5h-5k.However, unlike the synthetic data, here it can be seen thatno cda has more higher frequency components that real data,which can also be seen by the high variation over time in Fig-ure 5f. On the other hand, no ob has less higher frequency (oreven lower frequency) components which results in the flatshape in Figure 5g. The Stock-GAN spectral density, whileclosest to real one among all alternatives, also misses out onsome low frequency components. Nonetheless, Stock-GANis closest to real data due to our novel structural approach ofthe CDA network and use of order-book data.
We showed the superior performance of Stock-GAN in pro-ducing realistic market order streams compared to other ap-proaches. In doing so, we also introduced five statistics tomeasure the realism of generated stock market order stream.We chose our real GOOG data for dates in which there wereno external events, such as financial performance report.Thus, we did not model the effect of exogenous factors onstock market, which we believe is technically possible by justadding another condition for the generator. Notwithstandingthese effects, we demonstrated that stock market data canbe generated with high fidelity which provides a means forconducting research on sensitive stock market data withoutaccess to the real data. In future work, we intend to test theeffectiveness of the Stock-GAN on more stocks, other thanPN and GOOG that we did in this work.
Acknowledgement
We thank OneMarketData for the dataset used for this work.Most of this work was done when Junyi, Yaoyang, andArunesh were at the University of Michigan, where the workwas supported by US National Science Foundation undergrant IIS-1741190. eferences [Abe and Nakayama 2018] Abe, M., and Nakayama, H. 2018.Deep learning for forecasting stock returns in the cross-section. In
Pacific-Asia Conference on Knowledge Discoveryand Data Mining , 273–284.[Abergel et al. 2016] Abergel, F.; Marouane, A.; Chakraborti,A.; Jedidi, A.; and Muni Toke, I. 2016.
Limit Order Books .Cambridge University Press.[Arjovsky, Chintala, and Bottou 2017] Arjovsky, M.; Chin-tala, S.; and Bottou, L. 2017. Wasserstein generative ad-versarial networks. In , 214–223.[Bao, Yue, and Rao 2017] Bao, W.; Yue, J.; and Rao, Y. 2017.A deep learning framework for financial time series usingstacked autoencoders and long-short term memory.
PLOSOne
Computer Vision and Image Understanding
Advances in Neural Information ProcessingSystems , 77–87.[Chung et al. 2015] Chung, J.; Kastner, K.; Dinh, L.; Goel,K.; Courville, A. C.; and Bengio, Y. 2015. A recurrent latentvariable model for sequential data. In
Advances in neuralinformation processing systems , 2980–2988.[Da Silva and Shi 2019] Da Silva, B., and Shi, S. S. 2019.Towards improved generalization in financial markets withsynthetic data generation. arXiv preprint arXiv:1906.03232 .[Friedman 1993] Friedman, D. 1993. The double auctionmarket institution: A survey.
The Double Auction MarketInstitutions, Theories and Evidence, Addison Wesley .[Goodfellow et al. 2014] Goodfellow, I.; Pouget-Abadie, J.;Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville,A.; and Bengio, Y. 2014. Generative adversarial nets. In
Advances in Neural Information Processing Systems , 2672–2680.[Gulrajani et al. 2017] Gulrajani, I.; Ahmed, F.; Arjovsky, M.;Dumoulin, V.; and Courville, A. C. 2017. Improved trainingof Wasserstein GANs. In
Advances in Neural InformationProcessing Systems , 5767–5777.[Hiransha et al. 2018] Hiransha, M.; Gopalakrishnan, E. A.;Menon, V. K.; and Soman, K. P. 2018. NSE stock marketprediction using deep-learning models.
Procedia ComputerScience arXiv preprint arXiv:1901.01751 .[Kumar, Biswas, and Sanyal 2018] Kumar, A.; Biswas, A.;and Sanyal, S. 2018. ecommercegan: A generative adversarialnetwork for e-commerce. arXiv preprint arXiv:1801.03244 . [LeBaron 2006] LeBaron, B. 2006. Agent-based compu-tational finance. In Tesfatsion, L., and Judd, K. L., eds.,
Handbook of Computational Economics . Elsevier.[Mirza and Osindero 2014] Mirza, M., and Osindero, S. 2014.Conditional generative adversarial nets. arXiv preprintarXiv:1411.1784 .[Neil, Pfeiffer, and Liu 2016] Neil, D.; Pfeiffer, M.; and Liu,S.-C. 2016. Phased lstm: Accelerating recurrent networktraining for long or event-based sequences. In
Advances inneural information processing systems , 3882–3890.[Press et al. 2017] Press, O.; Bar, A.; Bogin, B.; Berant, J.;and Wolf, L. 2017. Language generation with recurrentgenerative adversarial networks without pre-training. arXivpreprint arXiv:1706.01399 .[Qian 2017] Qian, X.-Y. 2017. Financial series prediction:Comparison between precision of time series models andmachine learning methods. arXiv preprint arXiv:1706.00948 .[Radford, Metz, and Chintala 2015] Radford, A.; Metz, L.;and Chintala, S. 2015. Unsupervised representation learn-ing with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 .[Shi et al. 2019] Shi, J.-C.; Yu, Y.; Da, Q.; Chen, S.-Y.; andZeng, A.-X. 2019. Virtual-taobao: Virtualizing real-worldonline retail environment for reinforcement learning. In
AAAI .[Wah, Wright, and Wellman 2017] Wah, E.; Wright, M.; andWellman, M. P. 2017. Welfare effects of market making incontinuous double auctions.
Journal of Artificial IntelligenceResearch , 541–547.[Wellman and Wah 2017] Wellman, M. P., and Wah, E. 2017.Strategic agent-based modeling of financial markets.
RussellSage Foundation Journal of the Social Sciences arXiv preprint arXiv:1907.06673 .[Xiao et al. 2017] Xiao, S.; Farajtabar, M.; Ye, X.; Yan, J.;Song, L.; and Zha, H. 2017. Wasserstein learning of deepgenerative point process models. In
Advances in NeuralInformation Processing Systems , 3247–3257.[Xiao et al. 2018] Xiao, S.; Xu, H.; Yan, J.; Farajtabar, M.;Yang, X.; Song, L.; and Zha, H. 2018. Learning conditionalgenerative models for temporal point processes. In .[Yu et al. 2017] Yu, R.; Zheng, S.; Anandkumar, A.; and Yue,Y. 2017. Long-term forecasting using tensor-train rnns. arXivpreprint arXiv:1711.00073 .[Zhang, Aggarwal, and Qi 2017] Zhang, L.; Aggarwal, C.;and Qi, G.-J. 2017. Stock price prediction via discover-ing multi-frequency trading patterns. In
Proceedings of the23rd ACM SIGKDD international conference on knowledgediscovery and data mining , 2141–2149. ACM.Zhang et al. 2017] Zhang, Y.; Gan, Z.; Fan, K.; Chen, Z.;Henao, R.; Shen, D.; and Carin, L. 2017. Adversar-ial feature matching for text generation. arXiv preprintarXiv:1706.03850 .[Zhang et al. 2019] Zhang, K.; Zhong, G.; Dong, J.; Wang,S.; and Wang, Y. 2019. Stock market prediction based ongenerative adversarial network.