There is a long tradition of research using computational intelligence, i.e. methods from artific... more There is a long tradition of research using computational intelligence, i.e. methods from artificial intelligence (AI) and machine learning (ML), to automatically discover, implement, and fine-tune strategies for autonomous adaptive automated trading in financial markets, with a sequence of research papers on this topic published at major AI conferences such as IJCAI and in prestigious journals such as Artificial Intelligence: we show evidence here that this strand of research has taken a number of methodological missteps and that actually some of the reportedly bestperforming public-domain AI/ML trading strategies can routinely be out-performed by extremely simple trading strategies that involve no AI or ML at all. The results that we highlight here could easily have been revealed at the time that the relevant key papers were published, more than a decade ago, but the accepted methodology at the time of those publications involved a somewhat minimal approach to experimental evaluation of trader-agents, making claims on the basis of a few thousand test-sessions of the trader-agent in a small number of market scenarios. In this paper we present results from exhaustive testing over wide ranges of parameter values, using parallel cloud-computing facilities, where we conduct millions of tests and thereby create much richer data from which firmer conclusions can be drawn. We show that the best public-domain AI/ML traders in the published literature can be routinely outperformed by a "sub-zerointelligence" trading strategy that at face value appears to be so simple as to be financially ruinous, but which interacts with the market in such a way that in practice it is more profitable than the well-known AI/ML strategies from the research literature. That such a simple strategy can outperform established AI/ML-based strategies is a sign that perhaps the AI/ML trading strategies were good answers to the wrong question.
We report successful results from using deep learning neural networks (DLNNs) to learn, purely by... more We report successful results from using deep learning neural networks (DLNNs) to learn, purely by observation, the behavior of profitable traders in an electronic market closely modelled on the limit-order-book (LOB) market mechanisms that are commonly found in the real-world global financial markets for equities (stocks & shares), currencies, bonds, commodities, and derivatives. Successful real human traders, and advanced automated algorithmic trading systems, learn from experience and adapt over time as market conditions change; our DLNN learns to copy this adaptive trading behavior. A novel aspect of our work is that we do not involve the conventional approach of attempting to predict time-series of prices of tradeable securities. Instead, we collect large volumes of training data by observing only the quotes issued by a successful sales-trader in the market, details of the orders that trader is executing, and the data available on the LOB (as would usually be provided by a centralized exchange) over the period that the trader is active. In this paper we demonstrate that suitably configured DLNNs can learn to replicate the trading behavior of a successful adaptive automated trader, an algorithmic system previously demonstrated to outperform human traders. We also demonstrate that DLNNs can learn to perform better (i.e., more profitably) than the trader that provided the training data. We believe that this is the first ever demonstration that DLNNs can successfully replicate a human-like, or super-human, adaptive trader operating in a realistic emulation of a real-world financial market. Our results can be considered as proof-of-concept that a DLNN could, in principle, observe the actions of a human trader in a real financial market and over time learn to trade equally as well as that human trader, and possibly better.
We describe a new public-domain open-source simulator of an electronic financial exchange, and of... more We describe a new public-domain open-source simulator of an electronic financial exchange, and of the traders that interact with the exchange, which is a truly distributed and cloud-native system that been designed to run on widely available commercial cloud-computing services, and in which various components can be placed in specified geographic regions around the world, thereby enabling the study of planetary-scale latencies in contemporary automated trading systems. The speed at which a trader can react to changes in the market is a key concern in current financial markets but is difficult to study latency issues using conventional market simulators, and is extremely difficult to study "in the wild" because of the financial and regulatory barriers to entry in conducting experimental work on real financial exchanges. Our simulator allows an exchange server to be launched in the cloud, specifying a particular geographic zone for the cloud hosting service; automated-trading clients which attach to the exchange can then also be launched in the cloud, in the same geographic zone and/or in different zones anywhere else on the planet, and those clients are then subject to the real-world latencies introduced by planetary-scale cloud communication interconnections. In this paper we describe the design and implementation of our simulator, called DBSE, which is based on a previous public-domain simulator, extended in ways that are partly inspired by the architecture of the real-world Jane Street Exchange. DBSE relies fundamentally on UDP and TCP network communications protocols and implements a subset of the FIX de facto standard protocol for financial information exchange. We show results from an example in which the exchange server is remotely launched on a cloud facility located in London (UK), with trader clients running in Ohio (USA) and Sydney (Australia). We close with discussion of how our simulator could be further used to study planetaryscale latency arbitrage in financial markets.
In seeking to explain aspects of real-world economies that defy easy understanding when analysed ... more In seeking to explain aspects of real-world economies that defy easy understanding when analysed via conventional means, Nobel Laureate Robert Shiller has since 2017 introduced and developed the idea of Narrative Economics, where observable economic factors such as the dynamics of prices in asset markets are explained largely as a consequence of the narratives (i.e., the stories) heard, told, and believed by participants in those markets. Shiller argues that otherwise irrational and difficult-to-explain behaviors, such as investors participating in highly volatile cryptocurrency markets, are best explained and understood in narrative terms: people invest because they believe, because they have a heartfelt opinion, about the future prospects of the asset, and they tell to themselves and others stories (narratives) about those beliefs and opinions. In this paper we describe what is, to the best of our knowledge, the first ever agent-based modelling platform that allows for the study of issues in narrative economics. We have created this by integrating and synthesizing research in two previously separate fields: opinion dynamics (OD), and agent-based computational economics (ACE) in the form of minimally-intelligent trader-agents operating in accurately modelled financial markets. We show here for the first time how long-established models in OD and in ACE can be brought together to enable the experimental study of issues in narrative economics, and we present initial results from our system. The program-code for our simulation platform has been released as freely-available open-source software on GitHub, to enable other researchers to replicate and extend our work.
For more than a decade Vytelingum's Adaptive-Aggressive (AA) algorithm has been recognized as the... more For more than a decade Vytelingum's Adaptive-Aggressive (AA) algorithm has been recognized as the best-performing automated auction-market trading-agent strategy currently known in the AI/Agents literature; in this paper, we demonstrate that it is in fact routinely outperformed by another algorithm when exhaustively tested across a sufficiently wide range of market scenarios. The novel step taken here is to use large-scale compute facilities to brute-force exhaustively evaluate AA in a variety of market environments based on those used for testing it in the original publications. Our results show that even in these simple environments AA is consistently outperformed by IBM's GDX algorithm, first published in 2002. We summarize here results from more than one million market simulation experiments, orders of magnitude more testing than was reported in the original publications that first introduced AA. A 2019 ICAART paper by Cliff claimed that AA's failings were revealed by testing it in more realistic experiments, with conditions closer to those found in real financial markets, but here we demonstrate that even in the simple experiment conditions that were used in the original AA papers, exhaustive testing shows AA to be outperformed by GDX. We close this paper with a discussion of the methodological implications of our work: any results from previous papers where any one trading algorithm is claimed to be superior to others on the basis of only a few thousand trials are probably best treated with some suspicion now. The rise of cloud computing means that the compute-power necessary to subject trading algorithms to millions of trials over a wide range of conditions is readily available at reasonable cost: we should make use of this; exhaustive testing such as is shown here should be the norm in future evaluations and comparisons of new trading algorithms.
The "ZIP" adaptive trading algorithm has been demonstrated to outperform human traders in experim... more The "ZIP" adaptive trading algorithm has been demonstrated to outperform human traders in experimental studies of continuous double auction (CDA) markets. The original ZIP algorithm requires the values of eight control parameters to be set correctly. A new extension of the ZIP algorithm, called ZIP60, requires the values of 60 parameters to be set correctly. ZIP60 is shown here to produce significantly better results than the original ZIP (called "ZIP8" hereafter), for negligable additional computational costs. A genetic algorithm (GA) is used to search the 60-dimensional ZIP60 parameter space, and it finds parameter vectors that yield ZIP60 traders with mean scores significantly better than those of ZIP8s. This paper shows that the optimizing evolutionary search works best when the GA itself controls the dimensionality of the search-space, so that the search commences in an 8-d space and thereafter the dimensionality of the search-space is gradually increased by the GA until it is exploring a 60-d space. Furthermore, the results from ZIP60 cast some doubt on prior ZIP8 results concerning the evolution of new 'hybrid' auction mechanisms that appeared to be better than the CDA.
IEEE Transactions on Evolutionary Computation, Feb 1, 2009
The "ZIP" adaptive automated trading algorithm has been demonstrated to outperform human traders ... more The "ZIP" adaptive automated trading algorithm has been demonstrated to outperform human traders in experimental studies of continuous double auction (CDA) markets populated by mixtures of human and "software robot" traders. Previous papers have shown that values of the eight parameters governing behavior of ZIP traders can be automatically optimized using a genetic algorithm (GA), and that markets populated by GA-optimized traders perform better than those populated by ZIP traders with manually-set parameter values. This paper introduces a more sophisticated version of the ZIP algorithm, called "ZIP60", which requires the values of 60 parameters to be set correctly. ZIP60 is shown here to produce significantly better results in comparison to the original ZIP algorithm (called "ZIP8" hereafter) when a GA is used to search the 60-dimensional parameter space. It is also demonstrated here that this works best when the GA itself has control over the dimensionality of the search-space, allowing evolution to guide the expansion of the search-space up from 8 parameters to 60 via intermediate steps. Principal component analysis of the best evolved ZIP60 parameter-sets establishes that no ZIP8 solutions are embedded in the 60-dimensional space. Moreover, some of the results and analysis presented here cast doubt on previously-published ZIP8 results concerning the evolution of new 'hybrid' auction mechanisms that appeared to be improvements on the CDA: it now seems likely that those results were actually consequences of the relative lack of sophistication in the original ZIP8 algorithm, because "hybrid" mechanisms occur much less frequently when ZIP60s are used.
Electronic Commerce Research and Applications, Jun 1, 2003
This paper describes the use of a genetic algorithm (GA) to find optimal parametervalues for trad... more This paper describes the use of a genetic algorithm (GA) to find optimal parametervalues for trading agents that operate in virtual online auction "e-marketplaces", where the rules of those marketplaces are also under simultaneous control of the GA. The aim is to use the GA to automatically design new mechanisms for agent-based e-marketplaces that are more efficient than online markets designed by (or populated by) humans. The space of possible auction-types explored by the GA includes the Continuous Double Auction (CDA) mechanism (as used in most of the world's financial exchanges), and also two purely one-sided mechanisms. Surprisingly, the GA did not always settle on the CDA as an optimum. Instead, novel hybrid auction mechanisms were evolved, which are unlike any existing market mechanisms. In this paper we show that, when the market supply and demand schedules undergo sudden "shock" changes partway through the evaluation process, two-sided hybrid market mechanisms can evolve which may be unlike any human-designed auction and yet may also be significantly more efficient than any human designed market mechanism.
automated mechanism design, auctions, marketplaces, ZIP traders GA, genetic algorithm, BICAS A se... more automated mechanism design, auctions, marketplaces, ZIP traders GA, genetic algorithm, BICAS A sequence of previous papers has demonstrated that a genetic algorithm (GA) can be used to automatically discover new optimal auction mechanisms for automated electronic marketplaces populated by software-agent traders. Significantly, the new auction mechanisms are often unlike traditional mechanisms designed by humans for human traders; rather, they are peculiar hybrid mixtures of established styles of mechanism. Qualitatively similar results (i.e., non-standard hybrid mechanism designs being evolved) have been demonstrated for Cliff's ZIP trader algorithm and also for Gode & Sunder's ZI-C traders, provoking the possibility that such hybrid markets may be optimal for any marketplace populated entirely by artificial trader-agents. The financial implications of this work could potentially be measured in billions of dollars. In an attempt to elucidate why these evolved hybrid markets outperform traditional human-designed mechanisms, this paper presents results from thousands of repetitions of the GA experiments. These data allow 2D projections of the 10-dimensional real-space fitness landscape to be made, which inter alia illustrate a surprisingly high sensitivity in the relationship between the fitness evaluation function and the resulting landscape.
algorithmic mechanism design, auction and negotiation technology, automated trading, ZIP traders,... more algorithmic mechanism design, auction and negotiation technology, automated trading, ZIP traders, genetic algorithms, e-marketplaces This paper builds on previous papers describing our ongoing research in automated market-mechanism design: using a genetic algorithm (GA) to find optimal parametersettings for software-agent traders that operate in virtual "e-marketplaces", where the rules of the marketplaces are also under simultaneous control of the GA. The aim is that the GA automatically designs new agent-based e-marketplaces that are more efficient than existing markets designed by (or populated by) humans. Das et al. recently demonstrated that ZIP software-agent traders consistently out-perform human traders in Continuous Double Auction (CDA) marketplaces similar to those used in the international financial markets. Cliff used a GA to explore a continuous space of ZIP-trader auction-market mechanisms, where the space of possible auction-types explored included the CDA and also two purely one-sided mechanisms. Surprisingly, the GA would sometimes settle on novel hybrid auction mechanisms partway between the CDA and a one-sided auction. Such results occurred when the market's supply and demand schedules were unchanging, and also when the schedules undergo a single sudden "shock" change halfway through the evaluation process. These results could prima facie support the hypothesis that hybrid auctions are in general preferable to the well-known CDA mechanism. In this paper we present new results that clarify that hypothesis. In our new experiments, more than one shock-change in supply/demand occurs during the mechanism-evaluation process; under this regimen, CDA auction-mechanisms are identified by the GA as optimal in three of the four experiments reported here, and in the fourth experiment a near-CDA hybrid mechanism was evolved. From these results we conclude that, while evolved hybrid market-mechanisms may be useful in niches where the marketplace's supply and demand dynamics are known a priori to be relatively stable or regular, the CDA remains the mechanism of choice when it is difficult or impossible to make accurate predictions concerning the dynamic stability of the supply and demand schedules.
algorithmic trading, online auction marketplaces, emarketplaces, automated market mechanism desig... more algorithmic trading, online auction marketplaces, emarketplaces, automated market mechanism design, trader-agents, ZIP traders, genetic algorithms The Zero-Intelligence Plus ("ZIP") adaptive automated trading algorithm has been demonstrated to outperform human traders in experimental studies of continuous double auction (CDA) markets populated by mixtures of human and "robot" traders . To successfully populate a market with ZIP traders, the values of eight control parameters need to be set correctly. While these eight values can be set manually, previous papers have demonstrated that values of those parameters can be automatically optimized using a genetic algorithm (GA), to tailor ZIP traders to particular markets, and also (by adding an additional real-valued numeric parameter) to automatically discover novel new forms of auction market mechanism that are more efficient than the CDA. This paper introduces a more sophisticated version of the ZIP algorithm, which is shown to produce significantly better results. The extended variant is known as "ZIP60", because it requires 60 real-valued control parameters to be set correctly, and the original ZIP algorithm is re-named "ZIP8" accordingly. Manually choosing the correct values for 60 control parameters would be a very laborious task, but it is demonstrated here that an appropriate automatic optimization process can discover good sets of values for the parameters. A simple GA operating in the 60dimensional parameter space is shown to produce ZIP60 traders with mean scores significantly improved over ZIP8s, but also with high variance in those improvements. A slight revision of this approach is shown to give results with even higher mean improvements and also with lower variance in those improvements. The revised approach involves giving the GA control over the dimensionality of the parameter space being searched, starting with an eight-dimensional space and then allowing the GA to automatically and gradually expand that space up to sixtydimensional only when the increased number of parameters leads to identifiably better solutions. The results from ZIP60, while better than ZIP8, show a greatly reduced incidence of cases where the GA discovers auction mechanisms that are significantly better than the fixed CDA mechanism. This may be interpretable as evidence that the earlier ZIP8 results (where improvements on the CDA were common) were consequences of the relative lack of sophistication in the ZIP8 algorithm.
Gode & Sunder's (1993) results from using "zerointelligence" (ZI) traders, that act randomly with... more Gode & Sunder's (1993) results from using "zerointelligence" (ZI) traders, that act randomly within a continuous double-auction (CDA) market, appear to imply that human-like convergence to the theoretical equilibrium price in such markets is determined more by market structure than by the intelligence of the traders in that market. This paper presents a mathematical analysis that predicts serious failures in ZI-trader CDA markets. The analytical predictions are confirmed by computer simulations. Thus, more than zero intelligence is required of trading agents to yield human-like CDA market behavior.
This paper describes the design, implementation, and successful use of the Bristol Stock Exchange... more This paper describes the design, implementation, and successful use of the Bristol Stock Exchange (BSE) a novel minimal simulation of a centralized financial market, based on a Limit Order Book (LOB) such as is commonly in major stock exchanges. Construction of BSE was motivated by the fact that most of the world's major financial markets have automated, with trading activity that previously was the responsibility of human traders now being performed by high-speed autonomous automated trading systems. Research aimed at understanding the dynamics of this new style of financial market is hampered by the fact that no operational realworld financial exchange is ever likely to allow experimental probing of that market while it is open and running live, forcing researchers to work primarily from time-series of past trading data. Similarly, universitylevel education of the engineers who can create nextgeneration automated-trading systems requires that they have hands-on learning experiences in a sufficiently realistic teaching environment. BSE as described here addresses both needs: it has been successfully used for teaching and research in a leading UK university since 2012, and the BSE program code is freely available as open-source on GitHub.
We describe a novel simulation of a contemporary realworld financial exchange: London Stock Excha... more We describe a novel simulation of a contemporary realworld financial exchange: London Stock Exchange (LSE) Turquoise, and we also introduce a newly-created adaptive automated trading strategy called ISHV, which exhibits realistic behavior in situations where large orders can radically shift prices before transactions occur. LSE Turquoise is a recently-introduced platform where buying and selling takes place on a pair of coupled trading pools: a lit pool that is visible to all traders; and a dark pool where large "block" orders are hidden from sight until they are automatically matched with a counterparty, after which the transaction is then revealed. Orders from traders are routed to the lit or dark pool depending on their size, and on the reputation of the trader issuing the order. Unlike all other public-domain adaptive trading strategies, ISHV can alter the prices it quotes in anticipation of adverse price changes that are likely to occur when orders for block-trades are publicly visible: so-called market impact. LSE Turquoise is intended to reduce the negative effects of market impact; something that we test with our simulator. We extend the existing BSE open-source exchange simulator to incorporate coupled lit and dark pools, naming the new system BSELD. We show ISHV exhibiting market impact in a lit-only pool, and discuss how a Turquoisestyle coupled dark pool reduces or eliminates that impact. We also show results from a Turquoise-style reputationtracking mechanism, which can be used for modulating trader access control to the dark pool.
Many of the world's major financial markets are electronic, in the sense that all communication a... more Many of the world's major financial markets are electronic, in the sense that all communication among traders and internal record-keeping at exchanges is entirely mediated and executed by digital computer systems and associated communications networks; and many such markets are also highly automated, in the sense that they are heavily populated by automatic algorithmic trading system which have largely replaced human traders at the point of execution in many spot markets. This has created significant demand for people skilled in writing and managing algorithmic trading systems. To provide a complete education and training in this field it is highly desirable to allow students/trainees to study the operation of their own algorithmic trading systems running live on a real financial exchange, interacting dynamically with other automated traders. This paper describes the Bristol Stock Exchange (BSE), a simulator designed and developed to meet that need. BSE provides a full implementation of the Limit Order Book (LOB) at the heart of modern financial exchanges, and includes reference implementations of several well-known leading algorithmic trading systems. BSE allows users to submit a variety of order-types including market, limit, fill-or-kill, timeto-live, immediate-or-cancel, iceberg; orders for specific actions at market-open and market-close; and linked pairs of contingent orders. BSE can be configured to allow empirical studies of issues in order routing between multiple exchanges and the performance of cross-market arbitrage trading algorithms. BSE also has provision for varying the exchange's fee structure, including implementing maker-taker and takermaker pricing models, The Python source-code for BSE, which has been under ongoing development and extension since 2012, along with extensive documentation, is freely available on the GitHub online public repository, and can be used as a publicdomain platform for teaching and research.
Proceedings of the 2002 Congress on Evolutionary Computation. CEC'02 (Cat. No.02TH8600)
algorithmic mechanism design, auction and negotiation technology, automated trading, ZIP traders,... more algorithmic mechanism design, auction and negotiation technology, automated trading, ZIP traders, genetic algorithms, e-marketplaces This paper builds on previous papers describing our ongoing research in automated market-mechanism design: using a genetic algorithm (GA) to find optimal parametersettings for software-agent traders that operate in virtual "e-marketplaces", where the rules of the marketplaces are also under simultaneous control of the GA. The aim is that the GA automatically designs new agent-based e-marketplaces that are more efficient than existing markets designed by (or populated by) humans. Das et al. recently demonstrated that ZIP software-agent traders consistently out-perform human traders in Continuous Double Auction (CDA) marketplaces similar to those used in the international financial markets. Cliff used a GA to explore a continuous space of ZIP-trader auction-market mechanisms, where the space of possible auction-types explored included the CDA and also two purely one-sided mechanisms. Surprisingly, the GA would sometimes settle on novel hybrid auction mechanisms partway between the CDA and a one-sided auction. Such results occurred when the market's supply and demand schedules were unchanging, and also when the schedules undergo a single sudden "shock" change halfway through the evaluation process. These results could prima facie support the hypothesis that hybrid auctions are in general preferable to the well-known CDA mechanism. In this paper we present new results that clarify that hypothesis. In our new experiments, more than one shock-change in supply/demand occurs during the mechanism-evaluation process; under this regimen, CDA auction-mechanisms are identified by the GA as optimal in three of the four experiments reported here, and in the fourth experiment a near-CDA hybrid mechanism was evolved. From these results we conclude that, while evolved hybrid market-mechanisms may be useful in niches where the marketplace's supply and demand dynamics are known a priori to be relatively stable or regular, the CDA remains the mechanism of choice when it is difficult or impossible to make accurate predictions concerning the dynamic stability of the supply and demand schedules.
Gode and Sunder 1993 described continuous double-auction cda markets populated by zerointelligenc... more Gode and Sunder 1993 described continuous double-auction cda markets populated by zerointelligence" zi traders that act randomly. Their results appear to indicate that no intelligence is required to give h uman-like trading performance. Cli and Bruten 1997 demonstrated serious failings of zi traders. Here, `zi-plus' zip traders are introduced: simple stochastic agents that adapt over time using an elementary form of machine learning. It is shown that the performance of zip traders is signi cantly closer to the human data than is that of zi traders. Thus, human-like trading behavior can be achieved with intelligence that is more than zero but much less than human.
Proceedings of the 8th annual conference on Genetic and evolutionary computation, 2006
The "ZIP" adaptive trading algorithm has been demonstrated to outperform human traders in experim... more The "ZIP" adaptive trading algorithm has been demonstrated to outperform human traders in experimental studies of continuous double auction (CDA) markets. The original ZIP algorithm requires the values of eight control parameters to be set correctly. A new extension of the ZIP algorithm, called ZIP60, requires the values of 60 parameters to be set correctly. ZIP60 is shown here to produce significantly better results than the original ZIP (called "ZIP8" hereafter). A genetic algorithm (GA) is used to search the 60-dimensional ZIP60 parameter space, and it finds parameter vectors that yield ZIP60 traders with mean scores significantly better than those of ZIP8s. This paper shows that this optimizing evolutionary search works best when the GA itself controls the dimensionality of the search-space, so that the search commences in an 8-d space and thereafter the dimensionality of the searchspace is gradually increased by the GA until it is exploring a 60-d space. Furthermore, the results from ZIP60 cast some doubt on prior ZIP8 results concerning the evolution of new 'hybrid' auction mechanisms that appeared to be better than the CDA.
Gode and Sunder's (1993) results from using "zero-intelligence" (zi) traders, that ... more Gode and Sunder's (1993) results from using "zero-intelligence" (zi) traders, that act randomly within a structured market, appear to imply that convergence to the theoretical equilibrium price in continuous double-auction markets is determined more by market structure than by the intelligence of the traders in that market. However, it is demonstrated here that the average transaction prices of zi traders can vary significantly from the theoretical equilibrium value when the market supply and demand are asymmetric, and that the degree of difference from equilibrium is predictable from a priori probabilistic analysis. In this sense, it is shown here that Gode and Sunder's results are artefacts of their experimental regime. Following this, `zero-intelligence-plus' (zip) traders are introduced: like zi traders, these simple agents make stochastic bids. Unlike zi traders, they employ an elementary form of machine learning. Groups of zip traders interacting in exper...
Uploads
Papers by Dave Cliff