Market data feeds: calculate the cost of every tick

Server racks used to process market data feeds and tick updates
Photo by panumas nikhomkhai on Pexels

What does one market data tick cost us? More than the vendor charge divided by the message count.

A useful measure includes licensing, transport, processing and delivery, divided by the updates that reach the consumers that can act on them. Run the measurement across one busy session this week before changing feed capacity or vendor coverage.

The raw tick count is a poor denominator. A duplicate, a recovery message and an update for an unused symbol all consume resources, but they do not create equal operational value. Start by defining the unit: for example, one valid, sequenced update for an entitled instrument, normalised and available to a named strategy or client-facing service.

Define the tick before pricing it

Different teams use "tick" to mean a trade, a quote change, a book-level update or a packet. Write down which one appears in each dashboard and invoice reconciliation.

A packet may contain several updates; one book update may change multiple price levels. Treating these as interchangeable makes comparisons between feeds and venues unreliable.

For the measurement, separate three counts:

  • Received: messages or packets observed at the network interface, including duplicates and recovery traffic.
  • Accepted: valid, in-sequence updates after decoding, gap checks and normalisation.
  • Useful: accepted updates routed to a consumer that is entitled to use them, after any intentional filtering or conflation.

Record each count by venue, channel, instrument group and consumer. Keep snapshot traffic and incremental updates separate. Snapshots establish state or repair a gap; they should not be counted as ordinary market changes. Likewise, count retransmissions separately. A feed with a low average message rate can still create a material recovery burst after packet loss.

Calculate at least two unit costs. The full operating cost per useful update is the allocated recurring feed, network, compute and support cost divided by useful updates in the period. The marginal cost is the additional infrastructure and service cost caused by handling and distributing more updates. The first informs sourcing and budgeting. The second helps with capacity planning. For a practical guide to the systems and providers involved, see Trading Technology's infrastructure coverage.

Measure the path, not just the vendor rate

Instrument the feed from the network interface to the final consumer. Capture packet and sequence counters at ingress, decode errors and recovery requests in the handler, then publish rate and latency at each fan-out boundary.

Use monotonic timestamps from a properly synchronised clock. Wall-clock timestamps alone can make small latency changes disappear inside clock error.

The ranges below are typical market ranges for production environments, not service-level guarantees. Protocol, hardware, venue load, co-location and measurement points can move the figures substantially. Use them as prompts for where to inspect, not as acceptance criteria.

MeasurementTypical market rangeWhat to check
Incremental message payload40-200 bytesWhether reported size includes transport headers, padding or bundled updates
Feed handler decode and normalise10-200 microseconds at p99Timestamp boundaries, CPU pinning, allocation and tail latency under bursts
Internal publish to consumer100 microseconds-2 milliseconds at p99Queue depth, fan-out count and whether consumers share a congested path
Gap recovery and replay100 milliseconds to several secondsGap frequency, replay volume and time until a consumer is safe to trade
Peak-to-average update rate3:1-10:1Capacity against open, close, news and auction bursts rather than daily mean

Do not infer latency from a single end-to-end number. A 500 microsecond p99 might be caused by decode, a full queue, a slow subscriber or timestamp placement.

Preserve per-stage histograms and correlate spikes with packet loss, retransmission, CPU saturation and consumer count. Track p50, p99 and p99.9; a mean can hide the tail that matters to automated execution.

Measure useful output as a rate as well as a total. Compare updates received with updates published, and record the reduction at every filter.

If the handler accepts 100,000 updates per second but a strategy subscribes to 2,000, the difference may be entirely legitimate. If the difference is unexplained, check symbol maps, entitlement filters, duplicate suppression and stale subscriptions.

Put licensing and delivery costs on the same sheet

Feed prices are not always simple per-message charges. Exchange and vendor terms may distinguish display from non-display use, professional from non-professional users, derived data, internal redistribution, external distribution and device or user counts.

A strategy consuming data can trigger a different entitlement from a screen showing the same values. Confirm the terms with the relevant vendor and exchange; do not assume that technical access grants every use.

Build a monthly cost ledger for each feed and use case. Include recurring vendor and exchange charges, connectivity, cross-connects, cabinet or cloud charges, handler compute, storage, monitoring and the staff time spent on operations.

Allocate shared components using a stated method, such as peak bandwidth or CPU time. Avoid allocating everything by symbol count if one busy instrument group drives most of the load.

Keep one-off implementation work separate from recurring cost. Include the ongoing burden of managing certificates, symbol changes, software upgrades, entitlement audits, incident response and recovery tests.

These costs may not appear on a data invoice, but they affect the cost of operating a feed. A short inventory of the platform and provider landscape is available in our trading technology coverage.

Run three views: cost per received update, cost per accepted update and cost per useful update. A wide gap between the first and third can indicate overbroad subscriptions, duplicate channels or unused data. It can also be the rational price of resilience, such as a redundant feed or full-depth book. Label the reason instead of treating every reduction as waste.

Run this checklist during the next session

Use a normal session and one known busy interval. Keep the sample window and instrument set consistent across feeds. Assign one engineer to counters and one operations or finance contact to the cost ledger, so that technical volumes can be reconciled against contractual use.

  1. Before the open: export current subscriptions, instruments, channels and consumer mappings. Mark which consumers require full depth, top of book, trades or snapshots.
  2. During the session: collect ingress bytes, packet rate, sequence gaps, duplicates, recovery volume, accepted updates and published updates in one-minute buckets. Preserve p99 and p99.9 latency by processing stage.
  3. At the busiest interval: record CPU per core, memory, queue depth, network utilisation and fan-out count. Check whether one slow subscriber backs up other consumers or whether the handler drops data under burst load.
  4. After the close: reconcile feed volumes and users against entitlements and invoices. Separate display, non-display and redistribution use where the contract requires it. Flag any unknown consumer rather than guessing its status.
  5. Before changing coverage: calculate the three unit costs and compare peak capacity, recovery behaviour and latency. Test any proposed filtering or conflation against the strategies and controls that depend on the updates.

Save the raw counters, configuration version and calculation assumptions. A unit-cost number without its denominator, sample window and allocation method is difficult to reproduce. Share the same definition with engineering, operations and finance before using it to justify a vendor change.

Turn the result into an operating decision

A lower cost per update is not automatically a better feed. Removing depth, widening conflation or dropping a redundant path can reduce spend while changing execution behaviour or recovery risk. Compare each option against explicit requirements: permitted use, instruments, peak throughput, loss recovery, latency distribution and the time needed to restore a trusted book.

Look first for changes that preserve those requirements. Remove stale subscriptions, correct duplicate routing, right-size handler capacity and identify consumers receiving data they do not use. Where licensed terms permit, test whether filtering can happen before expensive fan-out. Check that any derived or conflated output remains within contractual terms and that sequence and freshness indicators survive the transformation.

For any change, set a rollback threshold before deployment. Examples include a rise in p99 latency, a missed recovery deadline, a sequence gap that is not repaired, or a mismatch between the published book and a trusted reference.

Keep a representative replay or capture so the team can compare old and new behaviour, including burst periods. The aim is a defensible cost measure that does not conceal a reliability trade-off.

Count the updates your consumers can use, then price the path that delivers them.

Finish the week with a one-page result: unit definitions, sample period, three counts, recurring cost allocation, peak load, latency percentiles and unresolved entitlement questions. Add the owner and next review date. That turns a one-off exercise into a baseline for capacity planning and contract review.

Before closing the ticket, check one number yourself: what percentage of updates received during the busiest five-minute window reached an entitled, active consumer?

Frequently Asked Questions

How do you calculate the cost of a market data tick?

Add the allocated recurring costs for licensing, connectivity, compute and operations, then divide by a clearly defined count of useful updates. Keep separate figures for received, accepted and consumer-delivered updates so filtering and recovery traffic do not distort the result.

What counts as a market data tick?

It depends on the feed and the measurement, and can mean a trade, quote change, book-level update or packet. Define the unit explicitly and do not compare counts that use different definitions.

Which market data feed metrics should a broker monitor?

Monitor ingress rate, sequence gaps, duplicates, recovery volume, decode errors, consumer publish rate and latency percentiles. Track queue depth, CPU and packet loss alongside those counters to identify where delays or drops occur.

How often should a trading firm review market data entitlements?

Review them when adding a feed, changing a consumer or extending use to display, non-display or redistribution. A regular reconciliation against active users and services can also reveal stale subscriptions and unrecognised use.

Why is average feed latency not enough?

An average can hide brief but material delays during bursts or recovery. Compare p50, p99 and p99.9 latency by processing stage, and correlate tail changes with packet loss, queue depth and consumer behaviour.