How accurate is real-time market data? A 2026 benchmark of 158,077 quotes
US stock quotes sat within one basis point of the national best bid and offer in 92.7% of samples. In forex, the feed sat 0.13 bp from the consensus of two independent references that were themselves 0.69 bp apart.
On this page
- Market data benchmark 2026: the results in one table
- What does accurate mean when there is no official price?
- US stocks: 92.7% of samples within one basis point
- Forex: 0.13 bp, and cross rates that have to agree
- Crypto: a 3.15 bp gap that is really one number
- How often does a real-time price update?
- Market data latency: the number one clock gets wrong
- Does REST return what the stream just sent?
- Four tests you can run on any market data API
- What this benchmark does not tell you
- Questions
Key takeaways
- Across seven runs and 158,077 matched samples, 95% of US stock samples sat within 1.37 bp of an independent reference, and the median gap was 0.13 bp for forex and ETFs and 0.87 bp for perpetual futures.
- US stock midpoints were within 1 bp of the national best bid and offer in 92.7% of one-second samples, and within 5 bp in 99.65%.
- Outside US-listed stocks there is no official price. The two forex references sat a median 0.69 bp from each other, and the feed sat 0.13 bp from their consensus, closer than either reference.
- The raw crypto gap of 3.15 bp was mostly one constant: dollar stablecoins against bank dollars. With that 3.11 bp offset removed the tracking error was 0.81 bp.
- In these windows the median time between two quote messages on one symbol ran from 64 ms for ETFs to 305 ms for crypto, and the network round trip alone was 21 ms from London and 256 ms from Tokyo.
In liquid markets a real-time quote can be held to about one basis point of an independent reference, and you do not have to take a vendor’s word for it. In seven benchmark runs in September and October 2026, a TickerLayer quote sat within 1 bp of the US national best bid and offer in 92.7% of samples, a median 0.13 bp from a forex consensus and a median 0.87 bp from a perpetual futures consensus. One basis point is 0.01%: one cent on a 100 dollar share.
The less comfortable finding is about the yardstick. Outside US-listed stocks there is no official price to compare a feed with, and the independent sources we used as references do not agree with each other. In forex they were a median 0.69 bp apart, which puts each of them 0.35 bp from their own consensus, while the feed sat 0.13 bp from it. A benchmark that leaves that number out has left out its error bar.
Each run is an observer outside the platform holding an ordinary API key: one public WebSocket connection plus REST calls, the view a customer has. Once a second it compares the latest TickerLayer quote with the latest reference quote for the same symbol. The seven runs add up to 158,077 matched samples on 114 symbols, read from about 2.2 million messages without a single reconnect.
- 158,077matched one-second samples
- 114symbols in 7 asset classes
- 0.13 bpmedian gap, forex and ETFs
- 92.7%of US stock samples within 1 bp
Market data benchmark 2026: the results in one table
| Market | p95 gap | Within 1 bp |
|---|---|---|
| Forex | 0.48 bp | 98.5% |
| US-listed ETFs | 1.04 bp | 91.6% |
| US stocks | 1.37 bp | 92.7% |
| Indices | 1.76 bp | 59.7% |
| Perpetual futures | 2.05 bp | 54.3% |
| Commodities | 3.33 bp | 47.3% |
| Crypto | 4.91 bp | 7.3% |
95th percentile gap to an independent reference, by market
The stocks run also held four European and Asian names against a dealer’s quotes: 5,614 more samples, kept out of the US row. They sat much further away, a median 16.7 bp. That dealer quoted spreads of 27 to 46 bp where the feed showed 5 to 13 bp, so it is a loose yardstick, and the per-symbol rows are on the stocks benchmark page.
The ranking does not follow how fast a market is. It follows how well defined “the price” is. The three markets at the top are major forex pairs, which many dealers quote within a fraction of a pip of each other, and US stocks and ETFs, which have a single consolidated best bid and offer to be measured against. Crypto trades on hundreds of venues in two kinds of dollar, and it sits at the bottom until you account for that.
What does accurate mean when there is no official price?
For a US-listed stock there is a referee. The national best bid and offer (NBBO) is the highest bid and the lowest offer across all US trading venues at each instant, published on a consolidated feed. A benchmark can hold a quote against it state for state.
Forex, crypto and perpetual futures have no such thing. A benchmark has to build its reference, and say how. Ours are a consensus of two separately sourced forex streams, the median of up to three regulated crypto venues that trade against bank dollars, and the median of up to four large perpetual futures venues. None of them feeds the price being tested.
Those references do not agree with each other, and that disagreement is measurable:
| Market | Refs apart | Feed gap |
|---|---|---|
| Forex | 0.69 bp | 0.13 bp |
| Crypto | 2.00 bp | 0.81 bp |
| Perpetual futures | 3.69 bp | 0.87 bp |
Take the forex row. Two separately sourced FX streams were typically 0.69 bp apart, and the feed was 0.13 bp from their midpoint. Below roughly half a basis point, a forex benchmark is measuring its references as much as the feed.
US stocks: 92.7% of samples within one basis point
The stocks run compared 24 US names, among them US:KO, US:JPM and US:XOM, with the NBBO once a second through 25 minutes of the regular session on 10 September 2026. That is 35,640 samples.
The TickerLayer midpoint was within 1 bp of the NBBO midpoint in 92.7% of them and within 5 bp in 99.65%. Ninety-five percent of samples were within 1.37 bp, and 99% within 3.26 bp.
US stocks: share of samples by gap to the NBBO midpoint
Why not 100%? A one-cent book on a large stock flips several times a second. Two observers sampling the same second can land a cent apart while both are right. The useful comparison is the gap against the stock’s own bid-ask spread:
| Symbol | Spread | p95 gap |
|---|---|---|
US:KO | 1.13 bp | 1.13 bp |
US:JPM | 3.69 bp | 0.99 bp |
US:XOM | 4.27 bp | 2.14 bp |
US:DIS | 3.82 bp | 1.43 bp |
US:IBM | 10.48 bp | 0.42 bp |
US:CAT | 13.02 bp | 1.30 bp |
For 23 of the 24 names, the gap at the 95th percentile was no wider than the stock’s own bid-ask spread.
The 12 US-listed ETFs in the next run were as close at the midpoint: 91.6% of 14,220 samples within 1 bp, and 99.2% within 2 bp. Their spreads are tighter, a median 0.96 bp, so a single tick is a larger share of the spread.
Forex: 0.13 bp, and cross rates that have to agree
The forex run held 20 pairs against the two-stream consensus through a London morning, 08:45 to 09:10 UTC. The median gap was 0.13 bp, and 98.5% of 29,679 samples were within 1 bp. For EURUSD the median was 0.11 bp, about an eighth of a pip.
Forex also allows a test that needs no outside reference at all. EURUSD times USDJPY has to equal EURJPY, or the feed is quoting an arbitrage that does not exist. In the live run the median residual on that triangle was 0.12 bp, and the 99th percentile 0.61 bp. Seven of the eight triangles tested came in at a median 0.07 to 0.12 bp; the eighth, through EURCAD, was 0.90 bp.
This one you can reproduce on history and get the same numbers we did. The script below checks four triangles on hourly closes for the week of 9 to 13 November 2026:
"""Do cross rates reconcile? EURUSD x USDJPY against EURJPY, hour by hour."""
import json
import os
import statistics
import urllib.request
API = "https://api.tickerlayer.com"
TRIANGLES = [("EURUSD", "USDJPY", "EURJPY"), ("GBPUSD", "USDJPY", "GBPJPY"),
("AUDUSD", "USDJPY", "AUDJPY"), ("EURUSD", "USDCHF", "EURCHF")]
START, END = "2026-11-09", "2026-11-13" # one full trading week
def closes(symbol, span, unit):
out, offset = {}, 0
while offset is not None:
url = f"{API}/forex/agg/{symbol}/{span}/{unit}/{START}/{END}?limit=5000&sort=asc&offset={offset}"
req = urllib.request.Request(url, headers={"x-api-key": os.environ["TICKERLAYER_API_KEY"]})
with urllib.request.urlopen(req, timeout=30) as res:
body = json.load(res)
for bar in body["results"]:
out[bar["t"]] = bar["c"]
offset = body.get("next_offset")
return out
print(f"{'cross':<8}{'bars':>6}{'median':>8}{'p95':>7}{'max':>7} gap in bp")
for a, b, cross in TRIANGLES:
x, y, z = closes(a, 1, "hour"), closes(b, 1, "hour"), closes(cross, 1, "hour")
gaps = sorted(abs(x[t] * y[t] / z[t] - 1) * 10_000 for t in x.keys() & y.keys() & z.keys())
p95 = gaps[int(len(gaps) * 0.95) - 1]
print(f"{cross:<8}{len(gaps):>6}{statistics.median(gaps):>8.2f}{p95:>7.2f}{gaps[-1]:>7.2f}")cross bars median p95 max gap in bp
EURJPY 118 0.22 0.73 2.14
GBPJPY 118 0.25 0.86 1.48
AUDJPY 118 0.21 0.79 1.16
EURCHF 118 0.22 0.74 1.42Across 118 hourly bars each cross sat a median 0.21 to 0.25 bp from the product of its two legs, and the worst hour of the week was 2.14 bp. Run the same script on any forex API you are evaluating. A feed whose crosses drift several basis points from its own majors is stitching prices from sources that are out of step.
Crypto: a 3.15 bp gap that is really one number
Crypto has the worst raw figure in the table: a median 3.15 bp across 19 pairs, with only 7.3% of samples within 1 bp. The signed gap explains it. All 19 pairs sat above the reference, by 1.6 to 4.4 bp, and the median signed difference for the run was +3.11 bp.
The reference venues trade crypto against bank dollars. Much of the world’s deepest crypto liquidity is quoted in dollar stablecoins, which trade at a small premium or discount to the dollar itself. A feed that aggregates that liquidity carries the difference as a constant offset. Remove that single number and the tracking error is 0.81 bp, well inside the 2.00 bp that separated the lowest and the highest reference venue.
The 24 perpetual futures contracts, from BTCUSDT to gold as XAUUSDT, came in at a median 0.87 bp, with 99.9% of 34,491 samples within 5 bp. The composite midpoint sat between the lowest and the highest reference venue 76.9% of the time.
How often does a real-time price update?
Several times a second on a liquid symbol, which is more than a polling loop can follow. This is the median time between two consecutive quote messages on the same symbol in each benchmark window:
Median time between quote messages on one symbol
Quotes arrive in bursts, so the average rate is lower than the median gap suggests: about 10 quote messages a second per ETF, 4 per forex pair, crypto pair and perpetual contract, and 2 per stock. A REST loop that polls once a second sees about one of every two stock updates and one of every ten ETF updates, which is the practical case for a WebSocket over REST.
Volume adds up fast. The client subscribed to 24 perpetual contracts received about 480 quotes and trades a second. The 12 ETFs sent 179 a second, the 20 forex pairs 143.
Market data latency: the number one clock gets wrong
Latency is the time you received a message minus the timestamp inside it. That subtraction uses two clocks, and it is only as good as their agreement. If your clock runs 30 ms behind, every message looks 30 ms faster than it was.
Our crypto run shows the trap. Its median delivery time came out at minus 25 ms. A message cannot arrive before it was sent, so that number is clock offset, not speed. We lead with the tail percentiles, which a small offset barely moves, and withhold a figure when the offset is larger than the thing being measured.
| Leg, in ms | p50 | p95 |
|---|---|---|
| Market event to platform, stocks | 78 | 230 |
| Platform to client, stocks | < 50 | 78 |
| Platform to client, crypto | n/a | 280 |
Then comes the part no vendor controls: distance. These are network round trips from the platform’s origin to a probe in each city, before any software does anything:
Network round trip from the origin, milliseconds
- 21 msLondon
- 41 msIstanbul
- 84 msNew York
- 140 msLos Angeles
- 184 msHong Kong
- 256 msTokyo
The same feed is 21 ms from London and 256 ms from Tokyo. Where your code runs decides more of your latency than any figure on a pricing page, so benchmark from the region you will deploy in, with a synchronised clock.
Does REST return what the stream just sent?
An application that reads a quote over REST and then follows the stream should not see the price jump between the two. Every ten seconds the observer requested a REST quote and compared it with the stream frames around the request. Of 825 REST quotes that could be paired with a stream frame, in six asset classes, 98.2% matched exactly and every one was within 1 bp.
The second consistency test is about history. The observer fetched the latest closed one-minute bar, then fetched the same minute from the historical endpoint at least three minutes later. Of 593 bars, 94.4% came back identical in open, high, low and close.
Six of the seven asset classes scored 97.8% to 100%. Commodities scored 65.6%, with late revisions to the close of up to 23.6 bp at the 95th percentile. It is the weakest number in the benchmark and it is published with the rest.
Four tests you can run on any market data API
- Gap to an independent referenceSample the feed and at least one source that does not feed it, once a second. Report the gap in basis points per symbol, absolute and signed, then take the median across symbols so one busy name cannot dominate.
- Update gapLog the time between consecutive quote changes on each symbol. Report the median and the longest silence during market hours. A price that repeats for 15 seconds on a major pair is stale.
- REST against the streamRequest a REST quote every ten seconds and compare it with the stream frames within a second or two. Count exact matches and the size of any difference.
- Refetch settled barsSave the last closed one-minute bar as you see it live. Fetch the same minute from history a few minutes later. The two should be identical; if not, your backtest and your live system are reading different data.
Measures the feed
- Comparing what each side showed at the same second
- Medians and 95th, 99th percentiles
- A reference the vendor does not consume
- A fixed list of symbols chosen before the run
- Signed and absolute gaps, both published
Measures something else
- Matching event N on one feed to event N on the other
- Averages, or “up to” figures
- A reference that is one of the vendor’s own inputs
- Only the symbols that scored well
- Latency from an unsynchronised laptop clock
What this benchmark does not tell you
- Each run is 20 to 25 minutes of one day. It describes observed behaviour in that window and is not a service level agreement.
- The windows were calm. A London morning and a mid-morning US session are the easy case. Bad ticks and disconnects cluster around data releases and market opens, which need their own test.
- The references are imperfect too. A dealer’s quote for a commodity or an index is indicative, and the forex and crypto references disagree among themselves by the amounts shown above.
- One observer, one region. Delivery times from Asia or the US west coast will differ by the network distances in the chart.
- Ten symbols are not in the tables. Of 124 symbols on the fixed lists, eight returned no comparison in their window, mostly Asian stocks whose market was closed, and two were excluded for a reference that did not match the instrument.
Those limits are also the reason to rerun it. A benchmark is worth reading when it has a run ID, a method version and a date on every figure, and when the uncomfortable rows stay in the table.
Questions
How accurate is real-time stock market data?
In a September 2026 benchmark of 35,640 one-second samples on 24 liquid US stocks, the quote midpoint was within 1 basis point of the national best bid and offer 92.7% of the time and within 5 basis points 99.65% of the time. One basis point is one cent on a 100 dollar share.
How do you benchmark a market data API?
Run four tests from outside with an ordinary API key: the gap to an independent reference in basis points, the time between quote updates, whether REST returns what the stream sent, and whether a closed one-minute bar is identical when fetched again later. Report medians and percentiles, not averages.
How often do real-time prices update?
On liquid symbols, several times a second. In the 2026 benchmark windows the median time between quote messages was 64 ms for US-listed ETFs, 200 ms for forex pairs, 238 ms for stocks, 265 ms for perpetual futures and 305 ms for crypto pairs.
What is a good latency for a market data API?
For an application that is not co-located, tens to low hundreds of milliseconds at the 95th percentile. Network distance sets the floor: about 21 ms round trip from London and 256 ms from Tokyo to a European origin. Always measure the 95th and 99th percentiles from the region where your code runs.
Why do two market data sources show different prices?
Outside US-listed stocks there is no single official price. In the benchmark, two independent forex streams were a median 0.69 basis points apart and the crypto reference venues 2.00 basis points apart at the same second. Stablecoin-quoted and dollar-quoted crypto prices differ by a further constant offset.
Is the price on REST the same as on WebSocket?
It should be. Of 825 REST quotes paired with a stream frame received within 1.5 seconds, in six asset classes, 98.2% matched exactly and every one was within 1 basis point.