One broker routes full depth from several venues into its own feed handler, accepting higher bandwidth and engineering costs to control sequencing and latency. Another uses a consolidated vendor feed, accepting some loss of venue-level detail in exchange for simpler operations and a smaller bill.
Both are common choices. Neither is automatically cheaper once the cost of a tick includes licensing, recovery, compute and the systems that consume the data.
The useful comparison is not a vendor quote against a network quote. It is the fully loaded cost of delivering a defined market update to a defined consumer, within a measured latency and completeness target. That gives an engineering team a basis for deciding which data to buy, which to process and where to spend capacity.
Set the denominator first
"Tick" is not a stable unit across providers. A packet may contain several book changes. One vendor may conflate updates, while a direct feed exposes every order event.
A normalized update should mean one change your application can act on, such as a change to a price level or a trade print. Record the source message count as well, so the conversion is visible.
Count updates by venue, instrument group and consumer. Separate quote changes, trades, depth events and status messages where the feed supports that distinction. For a multi-asset broker, a single blended rate can hide a quiet FX stream beside a bursty equities or futures book.
Measure both sustained load and bursts. A typical planning range for a busy venue feed might be 5,000 to 30,000 normalized updates per second in ordinary active periods, with bursts several times higher around opens, closes or major releases.
These are broad capacity-planning ranges, not a description of a particular venue. Replay captured peak periods when sizing handlers; a daily average will not reveal queue growth or packet loss.
Define the service delivered, too. A quote displayed in a client terminal, a price used by a risk check and an event delivered to an execution strategy may have different freshness and completeness requirements. Your internal service-level objective should state which consumer gets which data, and what degradation is acceptable.
Build the cost model
Use a monthly cost boundary that includes the production path and its resilience. The figures below are typical planning ranges in US dollars for a broker or prop firm deployment.
They are not exchange tariffs. Venue, product, non-display rights, user counts, region and contract terms can move licence costs well outside the bands, so confirm the applicable schedules before budgeting.
| Cost item | Typical range | Unit and scope |
|---|---|---|
| Venue or vendor data access | $1,000 to $30,000+ | Per venue or product group, monthly; rights and user fees can be additional |
| Feed handler and normalisation | $3,000 to $25,000 | Monthly software or managed-service allowance |
| Colocation and cross-connects | $1,000 to $8,000 | Per site, monthly; rack space and connectivity vary |
| Redundant network links | $500 to $5,000 | Monthly, depending on route diversity and bandwidth |
| Compute, memory and short-term capture | $1,000 to $10,000 | Monthly allowance for production and standby capacity |
Do not add those ranges mechanically and call the result a quote. Some vendor arrangements bundle access, normalisation and support.
Exchange charges may be billed by device, user, application or non-display use. A firm distributing data to clients also has reporting and entitlement work that an internal-only prop setup may not need. For a deeper view of the providers and systems in the stack, Trading Technology's infrastructure coverage can help frame the supplier landscape.
Calculate an operating cost per million normalized updates: monthly attributable cost divided by monthly updates in millions. For example, a $25,000 monthly stack processing an average of 20,000 updates per second for 6.5 trading hours a day across 22 days handles about 10.3 billion updates.
That is roughly $2.43 per million. The figure is not a universal price benchmark; it is a way to compare two architectures under the same workload and cost boundary. Run the calculation twice. Include all licence and resilience costs for the fully loaded view, then show marginal processing cost without fixed licence fees.
The first helps with sourcing decisions. The second shows whether additional symbols or consumers are expensive to serve once the feed is already paid for.
Measure the whole delivery path
A feed can be fast at the network edge and stale by the time a trading application reads it. Timestamp each useful stage: exchange or vendor timestamp where available, NIC arrival, handler decode, normalisation, internal publish and consumer receipt.
Keep clocks synchronised and document the clock source. If timestamps come from different clock domains, apparent latency can be measurement error.
Track distributions, not only averages. Record p50, p99 and p99.9 latency by venue and consumer, plus maximum queue depth and time spent behind a backlog.
A typical tuned local handler might target tens to a few hundred microseconds for decode and publish, but hardware, protocol and workload change that range. The target should come from the consumer's use case, not a generic latency number.
Useful operational measures include:
- Updates and packets per second, split by venue and message type.
- Sequence gaps, recovery requests, retransmitted packets and recovery duration.
- Handler CPU, memory, queue depth and dropped or rejected messages.
- Freshness at the consumer, including the age of the oldest update processed.
- Peak-to-average rate ratio and the duration of each burst.
Instrument recovery as part of normal capacity. A gap can trigger retransmission, snapshot processing and a temporary increase in CPU and network use.
Report time to detect, time to request recovery, time to rebuild the book and the number of downstream messages delayed or invalidated. A feed that recovers correctly but leaves a consumer stale for several seconds has not met the same service level as one that repairs in milliseconds.
Compare architectures on equal terms
Direct feeds offer venue-specific detail and control over sequencing, but create more protocol maintenance, connectivity and recovery paths. A consolidated feed can reduce the number of integrations and simplify symbol handling. It may also conflate updates, apply its own normalisation or introduce an extra processing hop. Measure whether those differences matter to each consumer rather than assuming the richest feed is always required.
Conflation is a trade-off, not simply data loss. A display may need the latest quote, not every intermediate change. A risk engine or strategy that depends on order-book transitions may need the full sequence.
Set policy by consumer: full depth and event sequencing for the applications that need it, a lower-rate or top-of-book stream for those that do not. Check licensing rights before redistributing or reusing a feed across internal systems.
For each candidate, replay the same captured session through the same downstream path. Compare update completeness, p99 and p99.9 delivery latency, recovery behaviour, peak resource use and cost per million updates.
Include a volatile interval and a gap-recovery test. A quiet-session benchmark gives a poor basis for a production choice.
Keep the comparison tied to the operational boundary. A cheaper feed can shift work into your own normaliser, symbol mapping and support rotation. A managed service can cost more on paper while reducing failure modes and on-call burden.
Record engineering hours and incident impact as separate costs if they cannot be priced reliably. The infrastructure discussions at Trading Technology also cover how platform and connectivity choices shape those boundaries.
Use a repeatable test before changing
Run a controlled baseline before negotiating a new feed or changing the handler. Capture at least one normal session and one high-volume session, preserve the raw messages, and label the hardware, software version and network route. A test that cannot be repeated will not settle an argument between procurement, operations and engineering.
Use this checklist for each feed and consumer combination:
- Define the unit. State what counts as a normalized update and how packet counts map to it.
- Scope the service. List venues, instruments, depth, users, consumers, rights and required freshness.
- Capture the baseline. Record normal and peak rates, latency percentiles, queue depth, CPU, memory and packet gaps.
- Test recovery. Inject or replay a sequence gap and measure repair time, backlog and consumer impact.
- Cost the full path. Include licences, handler, hosting, links, standby capacity and operational work.
- Replay candidates. Use identical sessions and downstream applications, then compare completeness and service levels.
- Set alert thresholds. Alert on stale data, sustained queue growth and recovery failure, not just process health.
- Recheck after change. Repeat the test after a venue, symbol set, software release or market-hour profile changes.
Review the model quarterly and after material changes in instrument coverage or client activity. A new consumer can increase non-display licensing exposure even when the message rate barely moves. A new venue can add modest steady traffic but a substantial burst and recovery burden. Keep those effects visible in the cost model rather than burying them in a general infrastructure line.
This advice does not apply unchanged to a firm whose strategy depends on proprietary exchange-level order events or a regulated record of every received message. In that case, completeness, auditability and venue-specific rights set the design boundary first; cost per tick is a secondary optimisation within it.
Frequently Asked Questions
How do you calculate the cost per tick?
Divide the monthly cost attributable to the feed path by the number of normalized updates received that month, expressed in millions. Calculate both a fully loaded figure and a marginal figure without fixed licence costs so sourcing and capacity decisions can be compared separately.
What metrics should a market data feed monitor?
Track update and packet rates, latency percentiles, sequence gaps, recovery duration, queue depth, CPU and memory. Measure freshness at the consuming application, since a healthy feed handler does not prove that downstream prices are current.
Are direct market data feeds always faster than vendor feeds?
No. Direct feeds can remove a vendor processing hop, but the result depends on route, handler implementation, hardware and recovery behaviour. Replay the same market session through both paths and compare end-to-end latency and completeness.
What makes a market data feed expensive?
The bill can include venue or vendor rights, non-display fees, feed handling, colocation, cross-connects, redundant links and recovery capacity. Licensing terms vary by venue, product, user and use, so confirm the relevant schedules rather than relying on a generic price.
Platforms · Software · Providers · Markets