A supply of granular information such as tick level data is essential to create sophisticated algorithms. However, most FX trading venues do not currently provide market data of sufficient depth, frequency or quality for really advanced modelling. This raises a number of questions including what efforts banks and trading platforms should be making to compile more comprehensive and improved data sets for use in Algorithmic FX applications. Some interbank platforms already sell historical tick data which can be utilised for back testing of strategies and research, which is encouraging. The question remains however, whether a lack of quality data will rapidly dim the current interest surrounding Algorithmic trading for FX. To debate the issue we invite: Jim OHagan, Product Manager, FX Market Data, ICAP, Yaacov Heidingsfeld, Chief Operating Officer at TraderTools, Justyn Trenner, CEO and Principal of ClientKnowledge, and Dmitry Bourtov who runs US-based Solaris Market Neutral Fund.

If one accepts the premise that FX data is by definition participant-specific with different data sets being required by both the buy-side and sell-side, what should be collected in real time? Every tick? Or just levels and aggregated amounts and what other factors are likely to be involved with making that decision?

OHagan: All professional market participants engaged in algo trading from both buy- and sell-side have an interest in collecting as much quality, granular data as possible; this is because strategies and models change over time, and data thats not used today may be required tomorrow. As proof of best execution becomes an increasingly ubiquitous and rigid requirement of doing business in FX, algo traders of every type will need access to highly granular, quality data. Collecting such data now will enable development of aggregated data sets and time series later.

Heidingsfeld: Before answering that question I think one should make an additional participant distinction; sell side versus buy side. There are many types of algorithmic trading not all of which will require the same data sets. However across the board we do not think individual ticks are necessarily helpful, but that levels with aggregated amounts are required information. Additional information such as; whether the market is trending, (ie are the spreads narrowing), are the spreads shifting, what is the dealable, executable depth of book, through one or more venues are all equally important as well.

Bourtov: To answer this question first we need to define what our Algo model is doing. If we are working on something long term which trades once per hour I dont think that you need to collect tick by tick data, usually people use some form of aggregation for such models. On the other hand if you are working on some sort of market maker strategy with average trade potential as one or two ticks you need much more detailed data like tick by tick or even DOM (Depth of market) data. Another aspect which we need to pay attention to is data collection versus data used to feed your Algo model. If your model does not require detailed data today it doesnt mean that you might not need it for other models tomorrow. So it is better to collect as much detailed data as possible transforming it to the required granularity to feed to your model.

Trenner: Looking at this from a sell-side perspective, there is a clear opportunity to systematize something that the sell-side has always done: the management of client flows and risk-taking or lay-off against your view of the opportunity inherent in the clients flows. That opportunity may be simply one of aggregation and internalization, or it may be that you believe you can enhance the opportunity by trading the information you gain by seeing the flow. Either approach is predicated on taking a view of the flow as having data value and the ability to manage your expectations and risk against that. It is something most do through traders gut feel. I have no doubt it is something you can do better by collecting every piece of client trade data accurately time-stamped and analyzing that against every tradable datapoint in the market. That is a straight data analysis proposition and so every banks client dataset is intrinsically unique.

The ability to collect all dealable data from all venues in a single place using smart order routing technology is now well established in equity markets where reference data is also dealable data, equal and accessible by all. This is not the case in the FX market. Does that mean that modelling FX algorithms will always be intrinsically different from modelling equity algorithms?

OHagan: The different regulatory environments in which the equity and foreign exchange markets operate essentially dictate the level of transparency in each of those markets. The equity markets are highly regulated and transparent; the FX market, on the other hand, needs to account for execution across a range of venues of varying characteristics. The data generated from each of these FX trading venues reflects that venues particular, unique characteristics resulting in the need to factor in many best execution risk scenarios. Feedback received from algo traders accustomed to designing programs for the equity markets, indicates that such models are not transferable to the FX space. In short, as long as the difference in transparency exists, modelling for equities and FX will be different.

Heidingsfeld: I think we should make sure we are all talking about algorithms the same way. We would define algorithmic trading as any trading where a computer is making the trade decision and sending out execution instruction to a venue where the order can be acted upon electronically. Included would be everything from very basic rules-based trading or smart order routing all the way up to fully automated pairs trading, statistical arbitrage, benchmark executions. Given the above set of premises, I think the answer is clearly yes, FX algorithims will always be different from those created to trade equities. Volume based algorithms are less meaningful in the FX world where volume and price are different for different participants. Different venues may be quoting different prices at the same time based on their own order book.

Bourtov: I think in general the answer is Yes. As far as FX data is different from equity or futures data based on various statistical benchmarks they might require different models to handle it. How far they are different are good questions. If you create something which makes decisions for the next month or so I dont think that logic in your model which trades FX will be very much different from an Equity model. On the other hand really short term stuff is very different. We also need to pay attention to the fact that the FX market as opposed to equities is very much distributed and it is hard to make decisions about real volume traded today versus yesterday. It creates a lot of problems for the models which used volume for decision making.

Trenner: Yes, for the reasons others have noted, the modeling will be different. However, it is clear that the need to have smart liquidity access and management (essentially the outbound messaging) fundamentally coupled with smart market data capture and analysis (the inbound side of the equation) is key. To this extent, the concepts are not so very different.