US$2 per million usable events is the cost in a simple example where a firm spends US$20,000 a month and processes 10 billion usable events. Divide that bill by raw messages instead, including duplicates and updates discarded by the strategy, and the apparent cost per tick falls. The infrastructure bill does not.
That gap is where many feed reviews go wrong. Teams compare vendor quotes or exchange fees, then treat the cheaper feed as the cheaper system.
A tick also consumes network capacity, parser time, queue space, storage and operational attention. The right denominator is the event that reaches the point of use with its meaning and timestamp intact.
Price the usable event, not the feed
Build a full cost numerator
Start with the monthly cost of delivering the data into the application that needs it. Include exchange and vendor charges, connectivity, cross-connects, appliances or cloud instances, egress, retention, monitoring and the engineering time spent maintaining adapters. Keep the figures separate before combining them. That makes it possible to see whether licensing or processing dominates.
Licensing is often the least portable line item. Fees and reporting obligations depend on venue, user type, display, non-display use, redistribution and the number of devices or applications. A feed that can be used for an internal pricing process may not carry the same rights for a client-facing display. Confirm the permitted use with the provider and exchange before treating two quotes as equivalent.
Choose a denominator that matches the job
Count events after the relevant filters and validation, then state what counts as usable. A market-making process may need every order-book change. A risk screen may need only trades and top-of-book updates.
For a chart, a bar or sampled quote may be sufficient. Do not compare those workloads by raw packet count. Track at least three rates: received messages, successfully decoded events and events consumed by the application. Record drops, duplicates and stale events separately.
A useful internal measure is total monthly cost divided by consumed events, reported alongside the cost per million received messages. That pairing shows both the bill and the efficiency of the pipeline.
A tick is not a packet
Protocol choices change the workload
One packet can contain several updates, and one book change can require several messages to reconstruct. Sequence numbers, retransmissions, snapshot recovery and channel failover all affect the work a handler must do.
A feed's advertised message rate may not equal packet rate, decoded event rate or the rate seen by a strategy. Define each unit before sizing the receiver.
Typical planning ranges vary by venue and message definition. The figures below are indicative operating ranges, not exchange specifications. Bursts can exceed them, especially at the open or during volatile periods. Normalised storage excludes database indexes and replication overhead.
| Feed workload | Typical hot-period event rate | Normalised storage per million events | Typical receive-to-handler latency |
|---|---|---|---|
| Top of book and trades, one active instrument | 1,000 to 20,000 events per second | 40 to 100 MB | 50 to 500 microseconds |
| Depth updates, one active instrument | 5,000 to 100,000 events per second | 60 to 160 MB | 100 microseconds to 1 millisecond |
| Venue-wide, multiple instruments | 100,000 to over 1 million events per second | 40 to 160 MB, depending on schema | 0.5 to 5 milliseconds in a shared or general-purpose stack |
These latency ranges describe the path from receipt to an internal handler, not exchange-to-strategy latency. A direct feed in a co-location facility and a vendor feed over a managed network are different products.
So are a tuned, pinned process and a shared service running beside unrelated workloads. Measure the path you operate, with the timestamps you can trust.
Normalisation can save work or add it
A common schema simplifies downstream code, replay and monitoring. It also adds parsing, mapping and sometimes copying. For one strategy, a native protocol handler may avoid a conversion stage and reduce latency. For several products, a shared normaliser can reduce duplicated adapters and simplify recovery. The decision turns on how many consumers need the data, not on a general preference for native or vendor formats.
Where the bill accumulates
Bandwidth is only the visible part of volume. A process can receive a manageable stream on average and still lose updates during bursts because its socket buffer, queue or CPU core cannot keep up. Backpressure that is hidden by a large queue can turn an overload into stale prices rather than an obvious drop. Monitor queue depth and event age, not just interface utilisation.
Storage compounds the same mistake. At 100,000 events per second, 100 bytes per normalised event produces about 864 GB a day before indexes, replicas and backups. Retaining every raw update for a year may be justified for audit, replay or research, but it should be an explicit retention decision. Separate the hot replay window from longer-term archival data, and test whether sampling or compression preserves the use case.
Cost reviews should distinguish fixed and marginal expense. An exchange connection or support contract may be a fixed monthly cost; an additional instrument can add licensing, compute and storage. A feed with a higher quote may need less engineering to reconcile symbol maps, timestamps and recovery.
Include failure handling and on-call work in the comparison. Coverage and recovery behaviour matter as much as the nominal update rate.
For broader platform context, Trading Technology's infrastructure coverage tracks the systems and providers that sit behind broker and prop-firm operations. For a specific feed decision, the useful evidence still comes from a measured test on your own path.
Test the path before changing providers
Run a representative workload through the complete route: receive, decode, normalise, distribute and consume. Use a volatile session or a recorded burst, not only a quiet midday sample.
Compare feeds on the same hardware and with the same subscription scope where possible. Record both steady-state performance and recovery after a disconnect or sequence gap.
Before signing off a change, capture these measures:
- Peak packets and decoded events per second, plus the duration and frequency of bursts.
- Median and tail latency from capture to consumer, with clock source and timestamp point documented.
- Queue depth, CPU use, memory use, packet loss, sequence gaps and recovery time.
- Monthly licensing and connectivity cost, with display and non-display permissions listed.
- Storage per day, retention period, replica count and replay performance.
Then compare the total monthly cost per usable million events, while keeping the latency and recovery results beside it. A single cost ratio cannot express the value of a faster or more resilient feed. It can expose a quote that looks cheap only because processing, storage or operational work sits elsewhere in the budget.
This week, instrument one production feed from packet capture to consuming application and calculate its cost per usable million events.
Frequently Asked Questions
How do you calculate the cost of a market data tick?
Add the monthly feed, licensing, connectivity, processing, storage and operational costs. Divide that total by the number of validated events consumed by the relevant application, and report the result alongside the cost per million received messages.
What affects market data feed latency?
Network route, feed protocol, packet handling, parsing, normalisation, queueing and application load all contribute. Measure from a defined timestamp at receipt to the point where the consumer can act on the event.
Should a broker store every tick?
Only if retention supports a defined requirement such as audit, replay or research. Set separate retention periods for fast replay and archival records, and account for indexes, replicas and backups when estimating storage.
Platforms · Software · Providers · Markets