Algorithmic Market Making Strategies.pdf

Algorithmic Market Making Strategies Sahand Haji Ali Ahmad, Ph.D

Market Maker’s business model • Traders obtain liquidity on a bi-lateral and bespoke basis from dealers • The products are often standardised (e.g. spot) but the OTC nature lies in the delivery of liquidity • The market maker’s model is based on ideally having a large symmetric flow providing him with little risk and a significant spread capture • The business has a feed forward as higher symmetric volumes and P&L out of spread capture, leads to Market Maker’s ability to lower the spread offered and as such to increase his market share while less liquidity and patchy toxic volume might lead to higher spreads and subsequently less trade& profit • P&L is formed of 2 components: Spread capture and Inventory P&L

Venues for Market Makers • For Fixed Income (FX/Sovereigns/Corps) Market Makers are either banks quoting directly through GUI and API , or are banks and Trading firms quoting on the ECNs • For Equities, the market makers usually make markets on Equity Exchanges(Nasdaq, NYSE, Turquoise Plato of LSE, Euronext, Six Swiss, ….) or through dark pools of liquidity while Big Blocks are sometimes traded through brokers through IOI(Indication of Interest) • For Futures, the major venues are ICE, CME/CBOT/NYMEX • Different ECNs/Exchanges have different characteristics such as order execution time(30ms-70ms), latency, data update/reporting frequency(e.g. 20ms for EBS Live Ultra), reported data, type of orders(Iceberg,…), limits on size of the orders, settlement, stock tick size, odd-lot trading, block-trading, daily price limit, fees, type of Auction(continuous/fixing), fragmentation of the market, … • In equity markets, due to introduction of MiFid2, less volume has to be traded on Dark pools, so more volume is flowing to Auctions on lit exchanges

Pricing • Mid price/Spread • A naïve basic market-making model will combine the prices at various ECNs (FX/FI ECNs such as EBS(Nex Markets) , Reuters/FXall, NordFx Integral, Hotspot, Fastmatch, Fxspotstream, Currenex, FX Connect, Tradeweb, NYSE Bonds, MarketAxess, Bloomberg MTF, Nasdaq espeed, ICAP’s ETC, ICE, LME, CME, ….. ) Based on a linearly weighted combination and use it as the midprice for the clients • Some price the mid as a random number based on a normal distribution, with mean equal to the previous tick midprice • Spread should be proportional to the risk levels (expected short-term volatility in prices) • Spread should encourage the clients to trade (should be competitive and set dynamically to invite hopefully offsetting two-directional volume) • Spread for bigger trade sizes should be higher (2,5,10,20,50 Mil) in order to compensate for the excessive risk transferred by these bigger trade sizes

Mid Price with Signal • Current Mid Price to client should be a weighted average of mid prices available in the market shifted by an amount equal to expected change of the price until the (expected) time the trade is exited ( average time of staying in a trade based on FIFO exit method) so we have • Skew signal (Skew Tick) • If based on current flow, the average exit time is 20 Seconds, a few predictive signals for 20 seconds price change should be used to predict the 20 second price change, which will be subsequently added to the current mid price derived from orderbooks • The signals are proprietary and can be based on orderbook, volume, trend, executed price,… and their interaction, as well as their temporal evolution

Mid Price with Signal • • • • • • • •

There are various information used to generate the signals: Price action & Technicals Volume (& it’s change) Orderbook state (& Change in state of the orderbook) News sentiment Tweet sentiment Post-event analysis Orderflow (Information in orderflow, based on various client categories and weighed and aggregated based on their historic trade outcome similar to Alpha Capture from Analysts’ reports,….) • Other Assets’ prices/volumes/sentiment/…. (Correlated Assets)

Short-term signals • Looking at milliseconds or even seconds, the most important price predictor signal is the orderbook as the price is defined directly based on the mid of best bid and ask • Different viewpoints on orderbook modelling: 1) Descriptive model, reproducing stylized facts (like Hawkes [Bacry et al., 2013], or [Robert and Rosenbaum, 2011] and [Fodra and Pham, 2013]) 2) Structural models, typically flow-driven models (see [Cont and De Larrard, 2013], [Huang et al., 2013] or [Smith et al., 2003]) 3) Theoretical models, like [Ro ¸su, 2009], [Bayraktar et al., 2007], [Abergel and Jedidi, 2011] or [Lachapelle et al., 2013].

A simple Liquidity Imbalance Signal

Market Participants in these equities

Participants and the imbalance at their trades

Observations regarding Imbalance signal Obvious observations: • HF market makers and HF prop Traders use market orders to consume liquidity on the weak side of the book(buying when imbalance is on average 0.6 and selling when it is on average -0.6) • HF market makers provide double-sided liquidity (through limit orders) when imbalance is less intense than -0.5 (consistent with the known fact that HF participant contributing to stabilize the price with their limit orders) • HFMM trade far more with limit orders (73%) than with market orders • Investment banks use more market orders than limit orders • Orderbook Imbalance is highly correlated with the rate of insertions and cancellations of limit orders near the mid price • as market makers, HFMM are expected to earn money by buying and selling when the mid price does not change much (relying on the bid-ask bounce). On the other hand, HFPT are typically alternating between intensive buy and sell phases which are based on price moves

Signal definition and stylized facts • Orderbook Imbalance is defined as: (Q(b)-Q(s))/ (Q(b)+Q(s)) just before the next trade occurs, where Q(b) is the quantity on the best bid and Q(s) is the quantity on the best sell

Signal strength • Given that these are large tick stocks, for smaller tick stocks (and FX) several price levels need to be aggregated in order to obtain the same level of prediction for future price moves • A strong predictive signal as the average mid price move after 10 trades as a function of the current imbalance

Mean-reversion of imbalance to zero after several trades

Possibility of imbalance sign change in case of significant imbalance • Strong imbalance may imply on a future price change, which in turn can create a depletion of the “weak side of the order-book which may cause an inversion of the imbalance, since the queue in second best price level of the order book, which is now “promoted” to be the first level, could be large. Regressing future imbalances on the imbalance just before the first trade:

Trade intensity based on Orderbook imbalance

• relationship between the imbalance signal and the trading speed/rate, we observe the imbalance-conditioned trading rate R+(in the direction of the imbalance) and R-(opposite direction of the imbalance) for each type of market participant, during all consecutive intervals of 10 minutes

Interpretation and Conclusion • Previous figures show unbiased estimators for the probability that a participant of a specific type trades in the direction (respectively, opposite direction) of the imbalance (or in other words, the relative speed of trading in the direction/opposite to the imbalance) • High Frequency Market Makers: the higher the imbalance in the orderbook, the less they trade. This effect does not seem to be related to the direction of their trades. It corresponds to an expected behaviour from market makers • High Frequency Proprietary Traders: the higher the imbalance, the more they trade in the similar direction, and the less they trade in the opposite direction • Institutional Brokers: do not seem to be influenced by the imbalance. They trade more with limit orders when the imbalance is intense, this may derive the price to move in the opposite direction

A simple twist of the signal • Signal value = The quantity at the best bid divided by the sum of the quantities at the best bid and asks. As underlined in the paper, this signal has more predictive power when the tick is large, in the sense the quantities at the best bid and ask describe better the short term liquidity in the orderbook on large ticks

Twisted signal • The relationship between the expected price move in the future 20 seconds and the imbalance is increasing. Sampled every update of the first bid and ask (quantity or price), the density is rather U-shaped. Major trading venues available are used and aggregated(Nyse, BATS, NASDAQ and Nyse Arca). We show expected 20-second P&L if the trade is done when the signal exceeds a threshold(as a function of the threshold). Saturation mechanism is visible.

Other potential signals • Intensity of market buy/sell orders ( average time between two market buy/sell orders) • Intensity of limit order arrivals • Average size of buy orders vs. average size of sell orders • Temporal Evolution of Orderbook across time • Cross-asset/correlation-based signals • Imbalance of Trades crossing the spread • Intensity of order modification/insertion/cancellation as well as it’s temporal behavior • Combining Various Layers of Orderbook in a smart way

The Methodology, evidence and tests needed to establish and prove a signal Tests helping establish a signal and prove its effectiveness: • observing the average price profile after the signal (e.g average price chart 100 trades after the signal is triggered) • Verifying statistical significance of the positive P&L for the trade triggered by the signal ( Absolute value of T-stat more than 2) • Regressing expected return for the trade over the signal value • Calculating correlation of the trade P&L and signal value How to set the parameter to optimize the signal performance (e.g for orderbook imbalance we look at the threshold that whenever imbalance exceeds that threshold , we enter a trade on that side passively/actively , the exit strategy or the window size we look to exit the trade,…) • Parameters ( threshold for trade, closing time,…) that optimize: 1. Sharpe ratio (expected return of the trade/volatility of returns) 2. Average P&L per trade

Intensity of the process providing equity liquidity at the first limit(0.5 tick away from the mid price)

Intensity of the process at the second limit (1.5 ticks away from mid price)

Intensity of process on the second queue given the length of the first queue

Intensity of the first buy based on the length of the queue on the first level sell

Price Change in Orderbook after News • Price changes either due to trade (clearing a level) or due to shift of quotes(cancelling and replacing at a different price) • Price is an aggregate of Book effect(quote shifts) and Deal effect(level clearance)

Spread • Spread exists due to price uncertainty, transaction costs, holding premium • Spread offered should be based on the base spread (based on the client classification), as well as a spread based on upcoming volatility/unpredictability in the prices, Market Liquidity, client’s credit rating, client’s latency, client’s trading behavior (spamming/latency arbitrage/… ), …. • Quoting Large Size Trades to clients: in order to quote a large size trade ( e.g. 100 Million USDJPY quote), the widening of the spread should be incremental • It could be based on a fixed rule, progressive based on the size, or be based on AI ( which previous quotes took the trade, what is the expected minimum quote for profitability,….) • Just before announcements ( such as NFP,…) spread is widened prohibitively by market maker and the spread gradually tightens as time passes and volatility subsides • Spread is also highly based on the tick size, exchanges with smaller tick size obviously attract more volume and liquidity as market makers can offer narrower spread on these exchanges • The effective spread is the actual difference between the bid and offer, including the direction of any price movements. Whereas the realized spread is found by taking average bids and offers over a period of time, and finding the difference between them

Spread determination • Higher turnover is easily achieved by reducing the operating spread – the difference between market maker’s own buy and sell quotes. However, reducing the operating spread also lowers income per trading cycle2 . To the contrary, increasing the operating spread increases income per cycle, but reduces turnover. There is a trade-off between spread and turnover, and the fundamental question is: what is the optimal operating spread, that will maximize market maker’s profit from a given security? • Optimal spread based on turnover : Dynamically changing spread in order to maximize profit ( expected spread * turnover) • Based on a simple formula: ds= s * volatility * dz for equities, one can conclude that: spread = 2 * s * vol * sqrt ( average time between two trades) • Based on option pricing theory, a spread is the price of a straddle, so : spread = 1.6 * s * vol * sqrt (average time between two trades) Where: average time between two trades = Average trade size/Volume over the interval • This means higher the volume and lower the volatility, then lower the fair spread!!! • In reality the coefficient is between 2.8-3.8 • Massive intraday seasonality as volume has intraday seasonality patterns (highest fx volume for a pair during market times for the pair and economic news release, highest equity volume at open/close/other markets open,…)

Spread based on Volatility • Changes in the business costs, generally attributed to order processing, inventory, cost of carry or adverse selection, should therefore have an important impact on the evolution of the spread over time but also the market competition is very important • In FX one can use a vector autoregressive framework to model changes to five minute indicative spreads using volatility, interest rate differentials and quote revisions which collectively proxy the independent effects of information arrival, adverse selection and inventory costs • The most important factor in determining spread is volatility which is usually predicted based on a Vector AutoRegressive model(e.g. GARCH,…) • Lower spread naturally increases volume , which potentially translates into higher profit (if it compensates for lower spread captured) • Lower spread increases risk as it makes the market maker’s price possibly the most competitive leading to a significant one-sided flow of orders specially before important events • In order to avoid risk, market makers significantly widen the price before volatile market moving events and during illiquid/low volume periods (e.g. EURNOK during Asian hours) • In sovereign bond markets, usually the spread is one tick size as the rates don’t change that often so bid and ask are one tick size apart

Making Market on ECNs • When making markets on ECNs the spreads are not as fixed as dealing to clients on API • Quoted spreads depend on a number of factors which are specific to each venue: rejection rule, and proportion of Latency Arbitragers • Spread can also be determined dynamically so as to maximize expected revenue( spread * volume) • Usually the designated market makers are subject to maintaining certain conditions, such as minimum quotation time, maximum allowed spread or minimum turnover • Last Look is also accepted on most of the ECNs • Volume is never available but EBS provides the number of tickets • Market Microstructure is crucial and has to be utilized optimally ( e.g. EBS Live Ultra updates every 20 ms, number of execution tickets,….)

Market Impact • Market impact is very important in: 1. Execution of a large order for allocation purposes 2. Execution of a large order for hedging purposes 3. Arbitraging on the market after potential large client trades (based on accurate evaluation of the market impact created by client trade) Much research available regarding market impact of Meta Trades and Child Trades for Equity but little available for FX. Empirical studies have shown that the influence of the market impact is transient, that is, it decays within a short time period after each trade Two main models of market impact are: • Gatheral, Schied and Slynko (GSS) framework in which the market impact is transient and strategies have a fuel constraint, i.e., orders are finished before a given date T • Cartea and Jaimungal (CJ) framework where the market impact is instantaneous and the fuel constraint on the strategies is replaced by a smooth terminal penalization There are other market impact models where there is a residual permanent market impact remaining after the execution

Some Market Impact results

Market Impact • Market Impact for Meta Orders (rakhlin,…) • Market Impact for Child Orders (Eisler) • Market Impact for both (Hawkes process, Lehalle,..) • Market impact for FX • Market Impact for Equity

Predicting Price move Predicting price move is critical for the following functions: • Pricing Mid (contributes to the price skew) • Pricing Spread ( Wider spread if price is predicted to move significantly) • Last Look ( trade refused if price move is predicted to be significant (predicting trade P&L until the average trade exit time e.g 20 seconds) but usually done within a window of 100-200 ms if a certain threshold in price move has been hit) • Hedging/Inventory Management (instead of passive/active hedging when inventory levels rise, being patient and holding on to Inventory if price move is predicted to be favorable to the inventory) • Execution (Very clearly short-term prediction of price move helps directly increase P&L while executing)

LeHalle’s paper • Trade arrival dynamics and quote imbalance in a limit order book • Impact of passive orders on the book ( every order even passive is expected to have implications on the price move for a while later, such as 20 seconds later , which should ideally be used in the prediction of price for 20 seconds later)

Iceberg Order Detection • Christensen papers regarding Iceberg detection on Globex

Pricing with Risk Management • Skew book/hedge (soft hedging) • Midprice is skewed in order to reduce the risk/hedge the position (a major long/short EURUSD position means skewing the price lower/higher) • Usually skewing starts when the position exceeds a certain threshold, it can also be asymmetrical • Skewing reduces the chance of adding to an already significant position and increases the probability of reducing the position • The relationship between skew and position size can be binary, linear or exponential based on the position sizing • Some HFT strategies are based on estimating Market Makers’s positions based on the price skew they offer • Average time between a risk-increasing trade and its offsetting risk-decreasing trade is called internalization horizon

Probability of risk reducing trades based on position and skew type Assume dealer adopts exponential price skewing and sets her risk limit so that the absolute risk position is within a $25mn corridor 95% of the time, i.e. the skewing threshold is R = $25mn/ 𝜑 −1 (97.5%) = $12.75mn. Based on a simple Poisson arrival model of trades with different intensities for buy and sell orders(assuming equal sizes), distribution of position(will be a stationary process in equilibrium) and horizon can be calculated for the equilibrium scenario

Average time of 19 seconds to internalize the position(exit the position based on FIFO method) for a major LP

Distribution and cumulative internalization times

Internalization time depends on the currency pair

Internalization in practice

Internalization horizon based on the trade size

Theoretical cost vs. risk-tolerance trade-off

Risk Management by Principal Dealers • Externalisation: Virtu Financial Inc (2014, p2) : “Our strategies are also designed to lock in returns through precise and nearly instantaneous hedging, as we seek to eliminate the price risk in any positions held.” • Internalisation: Bank of England, H.M. Treasury, and Financial Conduct Authority (2014, p. 59): “Market participants have indicated that some dealers with large enough market share can now internalise up to 90% of their client orders in major currency pairs”

Trade Internalization • Internalisation refers to the process whereby dealers seek to match staggered offsetting client flows on their own books instead of immediately hedging them in the inter-dealer market • Bank and non-bank liquidity providers running an internalisation model benefit from access to large volumes of order flow originating from a diverse set of clients. Rather than immediately offloading inventory risk accumulated from a customer trade via the inter-dealer market, flow internalizers may hold open inventory positions for a short time (often not more than a few minutes) before matching against the flow of another customer. By internalising trades this way, they can benefit from the bid-ask spread without taking much risk, as offsetting customer flows come in almost continuously • As non-bank Market Makers and liquidity providers such as XTX, Jump, …. have been doing more internalizing, this has contributed to less volume on EBS and Reuters • By internalising flows, banks seek to internally match (uncorrelated) flows originating from clients, branches, internal prop desks, and market making hedge positions. Previously such positions would either have been auto hedged through an external venue, left as residual market maker risk positions, or dealt on another platform (lost business)

Trade Internalization • For banks with large international branch networks, it’s not uncommon to find high volumes of low value commercial transactions ‘leaking’ out to competitor bank platforms. But, by funnelling such trades through to the internal matching engine, banks can plug the leaks, and provide better services to internal branches or external clients • At the other end of the spectrum, for banks servicing clients seeking to execute high value trades, such internalisation enables these trades to be executed with very little if any market impact, meaning that clients get ‘their amount’ done, with minimal market impact and information leakage, and less residual market making risk to the bank

Trade Internalization benefits • Higher profits through capturing a greater proportion of the bid-ask spread on FX flows • Lower brokerage costs, and reduced reliance on external liquidity pools • Deeper more consistent liquidity available to clients even in volatile conditions

Internalization stylized facts • Internalisation horizon: the length of time a given trade forms part of the LP’s risk position before it is fully offset by another trade in the opposite direction

Aggregator • Aggregator is the orderbook formed from pulling multiple liquidity pools together to create an aggregated orderbook • On many ECNs, such as EBS Direct, If one is a liquidity provider streaming into an aggregator, then it is possible that he doesn’t know what he is pricing. It is also potentially the case that he would not be clear on whether he would be pricing $1 million in total, or just $1 million out of a trade that is in total $50 million. Under the current industry-wide method, he cannot be sure that someone else is involved and may move the market • As an LP if one believes there are other LPs involved in the trade that will hedge quickly, then one is also forced to hedge quickly and will actually contribute to the move. These issues can be subdued by using internalization • Aggregation of one form or another is taking an increasingly prominent role across a number of over-the-counter markets. For example, the Dodd-Frank act mandates that trading of vanilla interest rate and credit default swaps now takes place on Swap Execution Facilities where a minimum of three LPs are required to compete for a trader’s flow(see Commodity Futures Trading Commission, 2013). Traders in the US Treasury and corporate bond markets adopt a similar approach where they request quotes from multiple LPs when they require liquidity

Hedging and Inventory Management(Holding Risk) • Naïve way of evaluating risk based on inventory position hitting certain thresholds • Evaluation of risk based on toxicity of the client flow • Evaluation of risk based on the correlation between various currency pair positions in the inventory and the VAR for the portfolio ( VAR-COVAR model) • Soft hedge(skewing the prices to the clients) • Hard hedge/execution (actively liquidating the position in the market if a certain threshold is hit) • client-based Hedging of the inventory(toxicity, signals,…): e.g. weighing the position in the inventory based on toxicity factor of the client flow • When there is little liquidity in a product (such as USDNOR during Asia hours, one would hedge using EURNOR trade such that NOR position gets cancelled out)

The most customized hedging • Holding on to the risk in a smart way is the key to monetize the flow • Ultimate customized way to manage the risk is hedging each client separately based on their trade track record combined with the other existing signals in the market , making a predictive model for their trade P&L and using it to proportionally hedging the client’s trades

TCA and Optimal Execution Strategies • Best execution requirements under the second Markets in Financial Instruments Directive • Standard execution methods include: V-TWAP/Target Close/Implementation Shortfall/Hunt/Value/momentum/steps • The more sophisticated execution strategies are based on Almgren-Chris optimization system (Price move risk minimization vs. market impact minimization) • Taking advantage of all the available venues and their liquidity patterns for the speediest optimal execution with the least market impact • Transaction Cost Analysis helps in understanding the cost of execution better which will improve Last look performance( improving delay/rejection ratio) • A P&L analysis of the child trades (as well as their risk contribution) is based on the FIFO model for the inventory

Optimal Execution Strategies • There are many execution strategies available but to find the optimal one, usually a Markov Decision Process modelling is done: • The objective is always to execute the trade as quick as possible subject to minimizing market impact/transaction cost • At the beginning of each epoch an Action is to be taken ( cancel the previous order(s) and send a new order(s) with a different size/price , do nothing and wait another time epoch for the current order(s), cancel and send market orders of different sizes,…) • Each time an order is posted, depending on the size, price distance from the top level, volume, volatility, Orderbook Imbalance, Spread, average time to fill an order based on its distance from the mid, trade imbalance,…… an expected fill ratio for the next epoch is assumed (with a known distribution) • Different ECNs have different liquidity patterns at different times of the day, so based on the passive/active order, there will be an according market impact • There is a penalty involved at each epoch depending on the size of the order to be executed (or penalty if the order hasn’t completed by the target time) • It is a Dynamic Programming/Optimal Control/Reinforcement Learning problem, solved numerically in a backward style

Various Execution Strategies • Implementation Shortfall • VWAP • ….

Client Flow Analysis • Analysis of client flow in terms of: 1. Client Reval (Decay Profile) : How quickly the profile becomes negative, is it consistent, ……. 2. Client Trades risk profile: Are trades mostly risk increasing/risk decreasing , ……. While some clients’ trades can’t be statistically significantly categorized, some are statistically significantly contrarian (bad market timing by corporate clients) , some are statistically predictive ( high frequency prop firms) , some are statistically risk decreasing (Prop firms,…) • Research shows with a high probability the clients which are statistically contrarian stay contrarian (corporate clients,….) , those who are statistically predictive remain predictive ( HFT , ..) , those without signal but with significant market impact remain so (Brokerage clients, ….), those with risk-decreasing profile remain so(HFT), …… • This fact means their flow can be optimally and reliably used for predicting the price and managing risk if studied and categorized smartly into various toxicity levels

Client Flow Analysis • Classifying client flow into various segments based on the Client’s trades’ track record, decay profile, risk profile,….. • Using this classification can help as a signal indicative of the price move, e.g Order Flow A is statistically predictive while order flow B has been statistically contrary for the price move • Pricing: clients with significantly toxic flow get quoted wider spreads(pulling liquidity), trades initiating from statistically Toxic/Contrarian clients can help improve pricing as they transfer information regarding price move direction which can be subsequently used for handling the inventory(or even reaching in the market and acquiring an inventory in anticipation) • Hedging/Inventory Management: Hedging clients in a customized way based on their decay profile ( if they create significant market impact shortly after, then holding the risk and managing it smartly is important) • Hedging can also be done in a smart way based on Client flow, where toxic clients, whom the trades have been losing against, can be classified into one stream and hedged aggressively while benign clients with poor market timing ability can be used as a stream where the risk is welcome on the book • E.g. Clients who tend to spam the markets and create major market impact on the markets, their flow gets classified within a stream with wider spreads, or kept on the book as a risk that needs to be held until the market impact is reversed

Trading with Inventory • Holding on to the risk (Inventory) in a smart way, hedging/double hedging the toxic flow, holding on to benign flow, predicting the price move, evaluating risk in a smarter way to have a clearer understanding of risk as well as P&L for the trade,… • Inventory P&L maximization needs a very delicate process, where flow information and market data are optimally processed, so as toxic trades are aggressively hedged/double hedged and benign flow remains in the inventory, so as when market move is predicted favorable to the inventory, inventory is held, but when it is predicted to be negative, market maker let go of the inventory as soon as possible or even double/triple hedge the trade (particularly useful for when the clients create major market impact, spam and spray their trades in multiple venues, and their market impact sometimes peaks at 20-25 pips 2.5 mins after their trade meaning market maker can’t profitably exit before 5-10 mins)

Last Look Strategies • There are some measures that market makers and ECNs can take to limit the exposure to latency arbitrage strategies(trading on stale indicative prices offered) and to market takers spamming the market. In FX, some ECNs allow liquidity providers ‘Last Look’: after a trader has traded on a market maker’s price then the ‘Last Look’ is a fixed period of time in which the market maker has an option to reject the trade. Generally the trade is rejected if in this fixed period of time the trade moves against the market maker beyond some threshold. The market maker is inferring that the trader may be taking advantage of the liquidity and is essentially withdrawing the price they made to market. Doing so can neutralize the effect of a latency arbitrage as well as providing protection against market spamming, at least over the interval of time that Last Look is active, typically measured in milliseconds. Market makers may also use Last Look trade rejections on price streams provided to traders, particularly for traders who trade at a higher frequency • last look comes in, because liquidity providers want to see if the market has moved or is moving or when they are assuming risk, which is all about making internalization work properly • LP’s streaming into client Aggregators are subject to adverse selection on their stale quotes which subsequently justifies use of Last Look strategies

Last Look Strategies • The purpose of Last Look is primarily to protect against trading on stale prices due to latency, and against certain trading behaviour. For instance, activities such as aggregation, order splitting or previous quote selection may result in more rejected trade requests. Therefore, the proportion of trade requests that are rejected due to Last Look will depend in part on the trading behaviour of the client and the platforms and connections through which the client trades. Also, Last Look rejects trade requests whenever the market price moves beyond the price tolerance in place, so other factors such as technical errors, pricing errors, and market moves may also cause trade requests to be rejected by Last Look. The objective is to protect the market maker from “Order splitting” or “Aggregation” • In its simplest form, last look specifies a tolerance level to an adverse price movement over the latency buffer which, when exceeded, makes the liquidity provider reject the deal request. I refer to this as onesided last look – others use the term “asymmetric last look” to highlight the feature that deal requests never get rejected due to a price move that is in the liquidity provider’s favour. Two-sided last look is a natural extension where deal requests are rejected on the basis of excessive price moves in either direction. This has loosely been referred to as “symmetric last look”, although a precise definition of what constitutes symmetry has been lacking. Last Look settings – including the prescribed time delay and the price tolerance – may vary by client, based on each client’s connection type, trading platform, trading pattern, …. . Classifiers/AI can be used to decide the time delay, price tolerance (threshold level) for clients • The objective is to minimize losses on client trades subject to a certain limit on his trade rejection ratio as such a classifier needs to be designed to flag those trades that the losses from them falls within the 5% (targeted rejection ratio) biggest losing trades of the client OR setting the threshold so as to reject any trades predicted to lose more than a certain amount • Since May 2017, the top 10 liquidity providers on bilateral, disclosed EBS Direct have cut their holding times from 93ms to 37ms, while reject ratios dropped from 5.3% to 3.15% mostly due to FX Global Code

Last Look Strategies • Often the decision is threshold-based , where if price moves from the stale price accepted by client by more than a threshold, the trade gets rejected. The threshold can conveniently be half the spread. • Smart Last Look policies can be based on using Classifiers/AI in order to detect what threshold makes sense depending on the client and market conditions. Predicting whether the trade will still be profitable given the stale price taken by client, based on market conditions, can be done using a smart classifier • Usual last look for the B2C and B2B dealing is 100ms, while various ECNs have their own rules • Trades can be rejected also due to limits on the amount of exposure that can be accumulated in relation to a particular counterparty over a defined period, credit checks, checking that the counterparty is not requesting to trade on non-permissioned currency pairs • Last Look is agnostic as to the causes of market movements and will reject a trade request whenever the market moves beyond the price tolerance in place during the prescribed time delay. The optimal time delay and price tolerance depend on many factors, such as connection latency of the client, his trade behaviour( market impact, latency arbitrage,….), spread for his stream,… • While some clients like certainty in price, others prefer certainty in execution. In these cases another type of last look is executed, where indicative price is quoted to the client and upon receiving the approval, the market maker executes the trade at the current price at the time of receiving the approval (subject to price being within a reasonable distance of the initially quoted price)

Pricing model for Last Look Analysis • The model used for analyzing last look is the market maker, being among N market makers quoting inside the client’s aggregator and client accepting the best price • The market maker price is modeled as a estimate of the true price plus an AR(1) measurement error (it’s noise assumed correlated among market makers)

Pricing and Hedging other assets • their differences(rates vs bonds, FX vs. equities, equities vs. corps,... Essentially how each of them has differences) , pricing is different cause FX is based on ECN prices of the same products ( + signals : Skew Book & Skew Tick) , but bonds and credit are illiquid ( and pricing is based on the available prices for Similar products , different maturity/duration , different bonds of same quality , ...) . Another point in pricing is deciding the spread based on volatility , risk , ..........Hedging is also different : backbone of hedging is using soft hedging ( skewing) based on threshold or VAR model, or hard hedging ( crossing the spread) , or crossing with other currencies ( hedging EURKRO in Asian time with taking the opposite position in USDKRO , so essentially eliminating the KRO exposure) , or in Bond market it is Duration hedging as much as possible , or in Corporate bond market , using bonds of different Durations for hedging, other corporate bonds ( same credit rating , same industry , .....) • Whatever I wrote for FX , I understand how it works for Sovereigns/Credit/Equity as well and create the slides

Treasury/Sovereigns market making • Venues : TradeWeb , Bloomberg’s FIT ( MTF) , EBS BrokerTec , Nasdaq’s espeed, Intercontinental bondpoint , MarketAxess , Liquidnet (Dark pool), MTS BondVision • Brokers and Execution Strategies are very important due to liquidity fragmentation : FlexTrade’s Algo Wheel • Major Market Makers: Banks , Trading firms ( Flow Traders, Bluefin , Citadel, IMC, Millenium Advisors, …) , Asset Management(Natixis, Advisors Asset Management, CIBC, ..) • There are major differences in pricing: • Prices of STIR and bonds have much lower resolution (e.g. price might change between two levels for an entire hour,….) and less variance. Relative to FX pricing is quite stale • Price change happens not often and if it happens it is either one tick up, or one tick down , so pricing model is tertiary where either the price next epoch is the same , a tick up or a tick down ( A classifier model needed) • A Markov model of pricing can be useful as the number of transitionary states are limited ( basically three states for majority of time, same price, one up, one down) where the states represent all the useful information and signals at the final state • Important Questions: • What is the next tick midprice? Up/down/unchanged? • What is the expected horizon to exit? (significantly longer than the 20 seconds for FX) • What is the expected price move by the exit time?

Pricing using buyside signals

Hedging Treasury/bonds • Hedging can also be based on Duration hedging (as the inventory of various bonds of different maturities has highly correlated assets) limiting duration to a small number close to zero • A more sophisticated way of risk evaluation/hedging is based on extracting principal components of yield curve, and setting risk evaluation based on thresholds on them , as well as hedging the portfolio trading bonds of other maturities to lower the components of PCA model of yield curve change close to zero • Also in many cases when the spread doesn’t change much (for example Italian bond Vs. German bond) one can hedge a less liquid one(e.g. Italian bond) with German bonds

Buy Side Treasury/Rate trading • Edith Mandell presentation/lectures • While pricing the bond based on orderbook and other signals, the yield curve could start to shape into exotic shapes • Based on the buy-side models of yield-curve (Edit Mandell stuff) this could create arbitrage opportunities which means the pricing should adjust to the expected shape for the curve • For market makers this might not be exploitable as they exit the position within a few minutes, where the prices and the curve haven’t changed much but buy side can exploit it by entering the position and exiting whenever the curve reverts back to a usual shape ( could be hours) similar to Statistical Arbitrage

Credit Market Making • Corporate bonds are often illiquid (anything off the run or not nearest maturity is often highly illiquid) • Hedging an illiquid bond in the market through active hedging is too costly • Easier way to hedge is to do Duration hedging with the more liquid bonds of the same name • Hedging can be done through hedging with CDS (CDS+bond=risk-free bond) • Hedging can also be done through similar bonds(in terms of industry and credit rating)

Credit Market Making Pricing • Pricing Credit products is very tricky as there is sometime little liquidity on ECNs to source the price from (A particular name and maturity bond might have last traded hours or days ago) • Pricing in the cases of illiquidity should be done through using Kalman Filter(Trumid, Benchmark Solutions,…) : • Pricing Model: A Linear model represented by a Kalman filter Why Kalman Filters are a good way to model the prices? • Provides a Linear relationship good enough and simple(For price change based on small-scale short-term moves one can ignore second-degree changes) • Other Classifiers can also be used , where based on the trade price of the related products between T and T*, the classifier predicts the price of the bond(or its yield) at time T* (a nonlinear model)

Pricing Credit Using Kalman Assuming the current time is T* and the last trade for that specific corporate bond/CDS ( e.g IBM 10-yr bond, 4% coupon bond or IBM 10yr CDS) was at time T, any transaction price between times T and T* for any RELEVANT product is used as Observables! • Observables: Transaction Price of Relevant(IBM other bonds indicative of credit spread for various maturities, IBM CDS spread various maturity, Treasury and rates products indicative of change in base rates) products traded after time T(those related products not traded in the interval can be assumed to have their last traded price) • Hidden State: Credit spread and yield at that moment for that specific bond ( e.g IBM 10-yr bond, 4% coupon bond)

Hedging Credit products • Hedging one illiquid corporate bonds (where there is no demand even if prices are skewed) through: • same name but other maturities ( duration matching) • CDS available on the same name, • similar bonds (industry/sector/size/rating) , …. • Match the credit exposure , and then hedge the IR exposure through bond positioning

Equity Market Making • Major market making activity on exchanges( NYSE,EuroNext , LSE, TSE, ….) by HFT Market Makers ( Virtu/KCG , Citadel, … ) • Also a lot of orders are aggregated at Brokers(Institutional such as Fidelity, Alliance Bernstein, … or Retail brokers such as TD Ameritrade, Etrade,… ) and the resulting flow is forwarded to likes of Virtu/Citadel/Hudson River Trading/GTS/Susquehanna/Two Sigma/… ) , or Banks (UBS, Citi ATD, Credit Suisse, Deutsche Bank, BNY Capital Markets, Cantor Fitzgerald, …) or traded on Exchanges (through the smart routers/optimal execution of the broker) or directed to Dark Pools • A lot of brokers internalize equity flow from clients • Significant activity through Dark pools (Auction-style of trading) • Significant activity through Institutional brokers for trading big blocks • Major competition around penny stocks where the tick size to the stock price is the highest ( highest spread/price) • Pricing based on the signal : relative value ( other similar: Industry, Sector, Size, market/cointegrated equities ) , momentum , mean-reversion , sentiment, orderbook imbalance, other orderbook signals,….. • Spread based on volatility • Continuous adjustment of midprice and spread to maximize P&L • Risk management by skewing prices (skew book)/taking off-setting position in similar equities ( industry/size/..) based on the position on hand in another equity ( so a good measure of risk of the book ,is not to look at the position in each equity separately but possibly look into the risk of the entire basket of similar stocks)

Equity OTC Market Making • Pricing big blocks by banks/brokers for clients/on the dark pools: pricing big blocks of equities that clients want to trade and executing it smartly on the exchanges/dark pools over a certain period ( Price wide enough to cover for the risk): Optimal execution Based on short-term equity signals, within an acceptable time-frame ( 1 hour/2 hours , …) Based on a good model for estimating Implementation shortfall ( within the execution horizon) • Quoted price to client is the estimated implementation shortfall + a margin of safety for taking risk (spread based on intraday trend direction/strength, volatility, liquidity, expected upcoming volatility ahead of news, releases,….) Based on the market conditions and all the Proprietary intraday signals ( volatility, volume , drift, news sentiment, news volume, the industry/sector , open gap, …) , a best estimation of execution cost is crucial for quoting the clients Nomura T-Cost model is a rough tentative model and a benchmark for estimating the market impact/implementation shortfall • Estimate cost of execution based on Nomura T-cost model and then executing better using proprietary intraday signals ( e.g. if the client sells his block position to the bank for 0.3% lower than the price currently traded on the exchange( 0.3% premium which depends on volatility, liquidity, expected volatility ahead of major releases/news , trend of the move,…) , and our model says the price will be higher in 1 hour (execution horizon), we try to execute majority of the position as close as possible to the 1-hour mark and so pocketing a profit) • Execution quality is extremely important( optimized if predictive intraday signals)

Intraday Equity Signals • Trend • Volume • Volatility • Real-time News/Tweet/blog sentiment • Options volume • Industry/sector trend • Relative value to the sector • Overbought/Oversold (Mean-reversion and other Technicals,…)

Transaction Cost Model • The Old model (Kein and Madhaven):

• D(Nasdaq) is an indicator either 1 or 0 • Trade Size is the ratio of the order value to market Cap • A number of models exist based on linear regression over factors such as Order size, Trade Volume, Bid-Ask spread, Stock Volatility, Exchange Structure,…..

Nomura Transaction Cost Model Impact 14.0 Post-Trade Period

Trade Period

10.5 7.0 3.5 0.0 0.0

0.1

0.2

0.3

Transient Impact

0.4

0.5

0.6 0.7 Time

0.8

0.9

Temporary Impact

Total impact over the trade period



(

1.0

1.1

1.2

Permanent Impact

)



(

METRIC =   S +   v   T +   v   T

)

2

• Where S is the average bid-ask spread,  is the volatility, v is the trade rate and T is the trade duration • Decomposes the cost into: Instantaneous Impact, Transient Impact, Permanent Impact

14

Usual Challenges to market makers • Sudden/Unanticipated market price change/volatility: Such as the outcome of CHF depegging or Brexit vote • Spamming:In order to reduce transaction costs some traders may choose to split up a large order into smaller standard size amounts and hit liquidity on multiple venues simultaneously creating massive market impact. This reduces the cost for the trader, but exposes liquidity providers to the risk that the market will run away from them as they try to exit this position. This activity is referred to as ‘spamming the market’ as The trader may also be accessing the same underlying source of liquidity on multiple venues if the best price on the ECNs is offered by the same provider

Usual Challenges to Market Makers • Latency Arbitrage: Trades originating from Latency arbitrage strategies( buy-side which takes advantage of connectivity latency differences between different market makers, and abuses the price provided having the largest latency) sometimes through aggregators picking up on stale quotes • Lack of transparency when streaming through Aggregators: A liquidity provider streaming into an aggregator, It is potentially the case that the LP would not be clear on whether he would be pricing $1 million in total, or just $1 million out of a trade that is in total $50 million. Under the current industry-wide method, one cannot be sure that someone else is involved and may move the market • Clients trading inside an Aggregator of LP’s price streams: Sometimes clients end up inadvertently picking up on the most stale pricing (sometimes the best price on their aggregator) hurting the market maker

Current Changes and Future Trends Market Microstructure is continuously evolving due to continuous changes in the rules and regulations as well as changes in ECN’s reporting, aggregation logic, matching logic, fees. Some examples of the current and upcoming changes in Algorithmic Trading for sell-side are: • MIFID2:MiFID II will change how asset managers search for dark liquidity, not only by closing down BCNs(Broker Crossing Networks) and limiting the amount of trading permitted in dark MTFs(Multilateral trading facility) below the directive’s ‘large-in-size’ (LIS) pre-trade transparency waiver, but also by placing more responsibility on the buyside for best execution and introducing greater separation between payment for research and execution services. These changes will expose poor execution quality to investors like never before, thus concentrating the minds of both brokers and buyside dealing desks. Buyside firms are taking their obligations under MiFID II very seriously, and some are viewing best execution as a source of competitive differentiation. Firms are reexamining their execution benchmarks, making greater use of TCA and examining venue toxicity levels ever more closely to navigate the dark liquidity landscape effectively ahead of MiFID II

Current Changes and Future Trends • MiFID II double volume caps for equities, where now one needs a respected trading mechanism that can match orders received above 100% of Large In Scale (LIS) thresholds determined per stock by ESMA, the European Securities and Markets Authority • MiFID 2 requirement of best execution service means increasingly more trades have to be handled by SOR(Smart Order Router) so there is an urge for developing optimal execution strategies on multiple lit exchanges(through Orderbook or Auction) and dark pools

Current Changes and Future Trends • Global FX Code(e.g. Principle 17 restricts last look) • EBS automatic single-ticket execution on the platform for orders of more than five million of base currency originating from manual traders, changing the fees for sweep, Select and single-ticket execution(instead of breaking a large order into smaller trades with separate tickets) • EBS providing new analytics such as cost of rejects, reject rates and last look to help improve internalisation and reduce costs • EBS showing whether the trade was part of a 50 Million dollar Trade or a 1 Million Dollar