Flow Imbalance Summary & FAQ

Summary

Q: What is Flow Imbalance (FI)?

It's a pricing skew technique used to drive more Spread PnL into the system.

It involves skewing the price such that the majority of clients pay more and subsidise the minority in a given time interval, whilst preserving the spreads displayed to clients.

The technique results in :

• No slippage.
• No additional visible spread.

This technique makes for more PnL being generated regardless of risk management technique used.

Image above, shows an aggregator mid (LPs all combined together) and our pricing model mid.

• First skew involves a little cluster of broker selling for higher
• Then sustained broker buying for cheaper (we skew our mid downwards meaning we are buying the majority at a discount.

History

Our pricing models operate with two different modes of optimisation

• Price Predictivity
• Years ago we trained our online machine learning to optimise for price predictivity. Does your skew anticipate where the price is going to go?
• This is still important, particularly for hedging and optimising your pricing.
• More recently we tweak the success function that the models were optimising for.
• Flow Imbalance Predictivity
• The imbalance of flow is much easier to predict. If you know in the next interval there will be a majority of buyers and a minority of sellers.

Make the MAJORITY pay more and discount the MINORITY

Example Calculation

Looking at a particular signal if you skew your price by that signal

• There will be some flow that has to pay less.
• There will be some flow that has to pay more.

Flow explorer tracks the balance of flow that had to pay more. It does so in units of equivalent USD Millions.

• Positive numbers represent a net successful prediction
• Negative numbers represent a negative prediction.

Feeds that are not using a predictivity mechanism typically hover around zero.

Real-time PnL Tracking

Mid Diff PnL We track the difference in spread PnL from filling against Mahi's Pricing model's mid rate and a configurable Benchmark mid rate. This is a spread neutral means of tracking and evaluation.

Comparative Skew

• Risk Inventory Skew: Consider the scenario, where retail sentiment has built up a large position in say USDJPY. Then they reverse their decision and want to go the other way. Risk inventory skew makes it cheaper for them to exit their position => Flow imbalance skew is more responsive to retail herding. Risk inventory skew is not very responsive.
• Aggregator Skew: Blending all LP feeds in a single aggregated rate, results in outsourcing your skew decision to your LPs, who in turn will be skewing according to their business objectives and not yours.

Retail Client Sensitivity

Applications

• Nearly all retail trading portfolios. Cross asset class.
• Market Making on Choice/Commission Free in Japan

FAQ

Q: Is retail flow skew sensitive?

Retail has very low skew sensitivity, low to medium sensitivity to slippage and high sensitivity to spreads.

By using techniques like flow imbalance it is possible to improve revenues without any slippage which in turn can allow for more competitive spreads a key factor in growth planning.

Q: Can we leave ourselves open for arbitrage by applying FI skew?

In practise we have to do so at levels that are imperceptible to users and for fungible instruments in sharp or aggregated environments at levels that don't produce an arbitrageable rate.

In softer environments even arbitrageable rates can be profitable as the soft retail flow is often skew insensitive. We have controls (NoArbitrageNode) to allow for both modes of operation.

Q: How much skew should we do?
With non-fungible instruments like CFDs we have gone with as high as 150% of spread. For FX we usually recommend keeping to 100% or below.

Q: Can we relate how much we skew to the other quotes we receive from LPs

As FI is delivered as part of our signals framework we can limit the amount of impact they have using the signal width limiting nodes

Q: Can we vary the amount of skew by timezone?
FI skew produces most value during the most liquid portions of the day. We sometimes run with a reduced skew amount during illiquid hours where spreads are wider to remain imperceptible to end users.

Q: What about position or risk or inventory skew?

Consider the scenario, where retail sentiment has built up a large position in say USDJPY. Then they reverse their decision and want to go the other way. Risk inventory skew makes it cheaper for them to exit their position.

Flow imbalance skew is more responsive to retail herding. Risk inventory skew is not very responsive

Q: What happens to brokered flow when applying FI?

They will receive both negative and positive slippage as it is brokered. We've had very little operational noise regarding this as we only broker flow that we can no longer monetise using other methods.

Even if we classify flow as "Broker" we often keep that flow within the book but slow them down via acute execution or liquidity reduction settings.

Q: What if there are operational problems introduced from applying FI?

When adopting our technology, we fully expect our clients to migrate to our systems and commercially validate our value-add figures as part of onboarding. They also only pay for the modules that are in use, allowing for managed rollouts. We can also gradually ramp up the amount of skew applied as part of roll out. Monitoring and validating client operational noise levels throughout.

Q: Are there any jurisdictional considerations?
Within Compass we can produce different pricing models for different regions and/or regulatory environments. So we can tune the amount of skew and limits around it via different pricing models.

We may also find that different jurisdictions have different Flow Imbalances. We are able to evaluate this as part of the Flow Imbalance Explorer. Does regional or global FI signals work best?

We can also produce separate pricing models or execution controls for "Retail" traders vs "Professional" traders.

Q: Can we tie in the classifier into Flow Imbalance?

Yes, we can attribute a higher weighting to flow that has already been identifiable via the classifier as informational.

Q: How can we measures it's impact?

We have found the best measure is to evaluate Mid Diff PnL

In terms of measuring PnL impact we've got a couple of approaches

• we observe our pricing model vs a benchmark and track the Mid Diff Pnl on skew coming in as trades occur
• we take your fill prices and shift those trade prices by a skew (applying the technique independently of price source). This is part of our Value Add Simulation framework