On-Demand Webinar Recording
TrueTick – Introduction to AI Ready Market Data
In this webinar, KX Senior Cloud Architect Mick Hittesdorf explains what AI-ready market data actually means and how to measure it.
This session is designed for market data and data engineering leads, quants and quant developers, data owners across trading, risk and compliance, and anyone responsible for data feeding AI/ML models.
Webinar Description:
In capital markets, every AI initiative rests on the same assumption: the data is right. Most firms only find out it isn't when a quant or trader spots the problem or gets a complaint. In this webinar, Mick Hittesdorf, product lead for TrueTick and Cloud Architect at KX, goes beyond "garbage in, garbage out." Mick explains what AI-ready market data actually means and exactly how to measure it.
Topics Include:
-
What "AI-ready" means for tick and time-series data, and why general-purpose data quality tools struggle with it
-
The six dimensions of data quality (timeliness, completeness, uniqueness, accuracy, validity, consistency) and the order to tackle them in
-
What happens to models and agents when data is late, incomplete or stale, and why point-in-time correctness matters
-
How to turn data quality into a score you can track, share with stakeholders and set SLOs against
-
A live look at TrueTick: tests written in Python and SQL in Jupyter, scheduling that knows exchange calendars, real-time Slack and email alerts, and a dashboard showing quality across every database
Watch the Recording:
More About Our Speaker:
Mick Hittesdorf | Senior Cloud Architect, KX
Mick Hittesdorf is a versatile, accomplished, leader with a passion for enabling clients to innovate and solve problems with data, analytics, and machine learning (faster and more efficiently with the power of the Cloud). Mick has led high-performing software and data engineering teams, managed global data science and analytics platform operations, developed low latency, proprietary trading systems, defined enterprise architecture strategies, and delivered advanced technology solutions.
Webinar Transcript:
Justin Fry: Thank you so much for joining us today. This is a webinar to introduce TrueTick.
Our speaker today is Mick Hittesdorf. Mick is the Global Lead for KX in Capital Markets and has many years of experience building, tuning, refining and delivering high-performance systems. If you have any questions during the webinar, please use the Q&A button in the middle of your screen.
We'll take questions at the end and very much look forward to interacting with everyone here. We appreciate all the questions sent in ahead of time; there were some very exacting ones, which Mick has been thinking about.
So Mick, over to you.
Presentation
Mick Hittesdorf: Thank you, Justin, for that introduction. Welcome, everybody, and thank you for your time today. My goal is to spend the next twenty to thirty minutes introducing you to TrueTick, and we'll leave some time at the end for questions.
As Justin said, my name is Mick Hittesdorf. For about the last year I've been spending a lot of time developing TrueTick, a new data quality assurance, observability, alerting and reporting product offered by KX. We just launched our private preview. If you'd like to learn more, and actually experience how to take advantage of the many capabilities of TrueTick that I'm going to discuss today, I'd encourage you to visit the TrueTick page on either the OneTick website or the KX website.
We'll provide that link at the end of the call. You can apply to join a select group of capital markets participants who want to improve their data quality and ultimately demonstrate that their data is trustworthy, in a systematic, transparent, metrics-driven way, so that users can rely on that data for the most important capital markets use cases. With that said, let's get right into the presentation. You're going to hear a lot about trust and trustworthy data, and I'll also touch on how that relates to AI-ready data.
If you leave with nothing else from the twenty minutes you spend here, it's this: your users, whether they're quants, traders or surveillance analysts, will not use data they don't trust. No matter how good your data platform is, and no matter what your vendors tell you about their data, its coverage, or how quickly they can deliver it on a real-time feed, once users discover defects in the data, their trust in it erodes very quickly. That's the problem we address with TrueTick.
What I've found at many places I've worked, and with many clients I've spoken with, is that data quality is often assumed until something breaks. And if you don't find the defects in your data first, your users inevitably will.
One common assumption, whether among data platform teams, market data managers, or the quants and traders who ultimately rely on the data for important research, modeling and trading decisions, is that the vendor has taken care of data quality. If your firm licenses data from a vendor, very often the vendor has made claims, or you have simply assumed, that the vendor is testing and validating the data and ensuring that it is consistently high quality.
But at very few firms that license data, or even ingest it directly from an exchange feed, is that actually verified. All along the chain there's an implied assumption that somebody is testing the data, validating it, and ensuring it's fit for purpose.
An assumption is not evidence. You may have heard the phrase "trust, but verify." Simply using a vendor's feed is not proof.
I've had this unfortunate experience personally. In a prior position, I was head of data engineering at a quantitative market-making firm. We were building a data platform, moving very quickly and having a lot of success, but we did not give sufficient attention to proactive data quality testing, validation, reporting and monitoring. We had a few issues. Only a few, but that's all it takes. If your traders or quants find problems before you do and have to tell you about them, that hurts your reputation and erodes trust, and by then it's too late, because the damage has been done. Trust is lost.
Bad data leads to bad outcomes: unreliable models, broken processes, and, if you're piloting or putting agents into production, hallucinations caused by missing, conflicting or contradictory data. Ultimately that leads to losses and institutional and reputational damage inside your firm. If you're on a market data team or a market data platform team, you've most likely lost much, if not all, of your users' trust. That's a bad place to be. That's the problem.
You don't have to take my word for it: data quality issues are in fact systemic and pervasive, but no one likes to talk about them, especially vendors who have made claims. I'm not trying to cast blame on any one vendor. Every market data vendor struggles with data quality.
It's a hard problem, given the volume and velocity of market data. But regulators and researchers have established that market data is often systematically wrong without users ever knowing it. Here's a specific example: a vendor made contractual claims that real-time data was being delivered within a certain SLA, but the historical delivery patterns showed that, in fact, it wasn't.
And it varied depending on the recipient and on certain circumstances of how the feed was being delivered. The promise was that data would be delivered in fractions of a second, but during peak market volumes, certain recipients were getting data that was several minutes old. Just by looking at a stream of ticks, that's not easy to detect. If you're not testing latency, you would never know.
And it's very possible that some of the market participants relying on that data made risk management, trading or compliance decisions that were, euphemistically, less than optimal.
There are other examples, such as those relevant to the Basel risk data aggregation principles (BCBS 239).
Only two of 31 banks were fully compliant with those principles, a decade after they were due.
About 13% of outstanding derivatives reported under EMIR have missing or outdated valuation data. And in another case, almost 10% of US monthly stock returns changed after a routine data update. So you can see how pervasive data quality issues are, and very often they're extremely difficult to detect unless you've made a commitment to systematic, proactive data quality testing.
So, no surprise, TrueTick is meant to address that problem. TrueTick is an enterprise-grade data quality testing and monitoring platform purpose-built for capital markets data. It takes a traditional software engineering best practice, testing and building test suites, which is very familiar to anyone who has been a developer. I have a software engineering as well as a data engineering background, and we've learned over decades how software quality improves.
You test your code on every commit, automate with CI/CD, run unit tests, system tests and integration tests, track test results, create tickets and fix bugs. That's how software quality improves, and that methodology is directly transferable to data.
So TrueTick adopts established software quality testing best practices and applies them to data quality testing and assurance, using many of the same tools, which I'll show you in a minute.
As I've framed the problem, your market data consumers, whether bond traders, risk managers or ML engineers, depend on market data. Undetected data quality issues erode trust, introduce errors and create risk. And traditional monitoring tools tend to operate at a level of granularity that doesn't allow sophisticated, rich, deep data quality rules and invariants to be expressed and tested.
Many tools will let you monitor whether a feed is up or down.
Very few will let you test complex relationships between different feeds, or within the same feed, or check against reference data such as symbology and calendars. They don't have the intelligence required to validate a market data feed, and that's what TrueTick provides. It's a capital markets data quality assurance and monitoring solution that leverages testing frameworks just about any modern developer is familiar with, such as Python, pytest and Jupyter Notebooks.
The solution converts your OneTick market data into a verified source of truth, and everyone benefits: data engineering and operations teams, consumers of the data, people building models, and your AI pilots, which absolutely depend on trustworthy, AI-ready market data to succeed.
I've talked a lot about data quality but haven't defined it. Data quality professionals have essentially reached consensus. The words differ slightly if you Google it or ask an AI like Claude what data quality is, but you measure data quality along these dimensions. They're incorporated into the DNA of TrueTick: every data quality test in TrueTick is classified into one of these six dimensions.
Timeliness: is the data available when you need it? Completeness: is all the data present? Uniqueness: is there duplicate data? Accuracy: how well does the data agree with an objective source of truth?
Validity: does the data conform to specific business rules or formats?
That's particularly important for the complex symbology you find in capital markets, such as ISINs, CUSIPs and RICs: validating that those are correct.
And consistency, which is very important both within and across feeds. For example, if you have a dataset or feed producing one-minute bars, the symbols in those bars should agree with your raw tick data. All the symbols in your raw tick data for the same time frame should be in the bars, and vice versa. Or you might have an internally developed security master, and you'll want to cross-reference the symbols in your feed against it.
With TrueTick, you can write consistency tests that ensure those rules and invariants are being met.
You'll see how pervasive these dimensions are in TrueTick and how they're used.
So why TrueTick? Fundamentally, TrueTick is a tool, a platform, but what it supports is a data quality improvement program: a commitment, every single day, to writing tests, collecting results, tracking results, watching trends, responding to alerts, fixing problems and writing new tests. That's how data quality progressively improves over time, and you need a tool like TrueTick to do it. It's the platform that enables a data-driven culture of continuous quality improvement. One way TrueTick enables that is daily data quality reports.
TrueTick sends out a report every day summarizing all of your T+1 (historical) tests as well as your real-time tests. TrueTick can monitor both historical and real-time feeds, broken out by database, schema and dimension. It provides an intuitive dashboard, which I'll show you in a minute, that lets you drill into issues, investigate alerts, see which tests ran, see the data the tests ran on, and see the test results.
I'm going to cover release one before I talk about release two, which is currently in development. On alerting:
TrueTick supports alerting via both Slack and email, so your operations teams are notified as soon as any test fails. TrueTick also includes a data catalog and AI integration, which I'll demo.
The data catalog accelerates data discovery and gives an LLM rich context. Claude Code is included with TrueTick, and Claude Code, together with the skills and MCP servers integrated within TrueTick, accelerates the authoring of data quality notebooks and tests. It becomes much easier to create new data quality jobs, onboard databases and write tests, and then to investigate failing tests with Claude, letting Claude look at the data and get the context it needs.
Claude can not only identify the root cause but also suggest remediations, for example doing a reload, or looking at an upstream, derived or parent source that may in fact be the source of the issue.
This is the last slide before the demo, and I want to give you a sense of the scope of the capabilities. As I mentioned, TrueTick does testing and validation, monitoring, alerting and reporting.
I won't spend a lot of time on the architecture, but TrueTick is a cloud-native Kubernetes application that we host in OneTick Cloud. We have various mechanisms to connect to your data sources. If you're running your data platform on AWS, we can connect to that. If you're on another cloud, there are options, for example Direct Connect and AWS Interconnect.
Coming soon is the ability to deploy TrueTick in what we call bring-your-own-cloud (BYOC) mode, the last option here, so if you want to run TrueTick within your own cloud account, we'll be able to support that as well.
Integrated scheduling is built in, so you don't need a third-party scheduler to schedule your jobs; it's a very intuitive way of scheduling.
It supports spot instances. TrueTick is very cost-effective because all jobs run on spot nodes, which lets you take advantage of savings of up to 80 or 90%.
Going back to testing and validation: as I'll show you in a minute, Jupyter Notebooks are used to author tests.
All tests feed into what's called a data quality score, which is how we make data quality objective. The score is the number of tests that don't fail, that is, tests with a status of pass or warn, divided by the total number of tests, computed at the database, schema and dimension level. I'll show that in the dashboard. It's compared against a configurable KPI, 95% by default, and you see red, yellow or green depending on whether your data quality SLAs are being met.
Both real-time and T+1 monitoring are supported, along with trend analysis. Alerting is supported via email and Slack, and daily data quality reports are sent out. I'll show you the web dashboard in just a minute.
There's logging, and there are ways to configure the recipients. So it's a very robust platform with a full set of capabilities, and more coming. This doesn't even touch on the AI integration I'll show you shortly. So let's go take a look at TrueTick.
Demo
Mick Hittesdorf: I'll show you the dashboard first, and how to find and investigate data quality issues, including a real-time alert. Then I'll show you TrueTick Studio, which is where you author tests and also where you see the notebooks and jobs that have run. TrueTick saves the state of every notebook that's rendered, so you have a full, high-fidelity record of each job, all the tests that were run, and the test output.
Then, briefly, I'll show you how Claude Code can be used within TrueTick Studio. Let's go to the data quality dashboard first, starting from the end and working backwards. What I'm showing you here is our internal OneTick Cloud dashboard, with the OneTick Cloud data that's currently onboarded to TrueTick.
We've been running an internal pilot of TrueTick for about six months now.
TrueTick actually got its start as an internal project for improving OneTick Cloud data quality. I was able to leverage the experience I mentioned at a previous employer to inform the architecture and capabilities of TrueTick. As I mentioned in the slides, we recognized that data quality is an industry-wide problem, and that this tool could be used by our customers to solve the same sorts of problems we're using TrueTick to solve internally in OneTick Cloud. What I'm showing you here is the summary view of the data quality dashboard.
It shows a ten-day data quality score for the entire OneTick Cloud data estate that I've loaded into TrueTick. We support over 250 venues and even more databases, since in many cases there are multiple databases per venue.
We've onboarded just a subset of those.
You can use this heat map view to see all the databases currently monitored by our internal TrueTick instance. You can see some of the composite equity databases, like the Australian, Canadian and EU composite equity databases. We have our indexes database, some futures databases, bars and daily databases. This shows you the data quality score by dimension. A score means one or more tests classified into that dimension have run, the results have been collected in TrueTick's database, and TrueTick has computed the data quality score. Here you can see 100%, certainly greater than 95.
There's also a configurable score threshold, in this case two points. So yellow means a score in the range of 95 to 97.
Anything above 97% is green. Anything below 95 is red, as you can see here. White is also informative: it means there's no test coverage for that dimension. For example, at a glance you can see the TDI index database doesn't have any timeliness tests, which is definitely a gap. At a minimum, you'd typically want to test that data is available to users at a specified time. Most firms want T+1 data before the open of the next day at the latest.
There may be much more aggressive requirements, such as within two hours after the close.
That's the sort of requirement a timeliness test can verify. So you can use this heat map to see, at a glance, the quality of your entire data estate, everything that has been onboarded.
This view shows your data quality trends over time. It aggregates both the T+1 and real-time feeds in OneTick Cloud. Except for this one day, which is a Sunday, when we run fewer tests (only 12 that day), everything looks good.
The score for that day was 75%, which doesn't meet the 95% KPI shown by this yellow line. But on all the other days, the data quality score exceeded the 95% KPI threshold, so it's green. You can look at many different time periods: thirty days, ninety days, and so on. I'll leave it at the last day and show you some features of the side panel. You can go into different databases, tables, tests and notebooks.
There's a rich set of filters you can use to zero in on data quality issues. Going back to this screen, these are the actual tests that ran. These are real-time tests checking latency on the US Comp database, and quote gap tests.
Real-time tests dominate the list here because, obviously, we run more of them. So let's say we just want to see our T+1 data quality.
That changes the view, with fewer tests. We see that data quality dips a little, but it's still above the 95% KPI.
You can come in here and see that there are some issues.
And you can investigate them. For example, this one says the symbol list for our PBI futures bars is incomplete. Click on this link and you can see the notebook that was run. If you want to see all the tests run in this notebook, there are 14 or 15 tests being run on this database. This is the idiomatic structure of a TrueTick data quality notebook. It includes the schedule.
All you need to do to schedule a notebook is add this cell and tag it with "schedule," and the scheduler will pick it up and run these tests. You can see all the output for the tests, and you can see here that there's a failure.
For example, there are some missing symbols.
For real-time monitoring, the process is very simple. Let's look at today. There's a timeline view. I'm going to remove this filter, go back to selecting all categories of tests,
and show you real-time for the last day.
What it shows is US Comp and the other real-time database; these are the two real-time databases in OneTick Cloud.
If I zoom in a little closer on today's real-time tests, every one of these bars represents a test.
And you can see that just after the open,
one of these tests failed: the MEMDB latency per mount test. If I filter to show just the failures, I can see the test that just failed. Again, I can click on the notebook and see the actual real-time test, and if I want to see all the tests in the notebook, there's the table of contents.
The nice thing about notebooks is you can combine charts and code. In this case, you get a visual view of the test. This is a latency test. A mount is a thread; there are multiple threads processing ticks for a real-time feed.
Per thread, this test validates that latency over a one-minute window does not exceed 1,000 milliseconds, or one second. You can visually see that this test failed. A handful of mounts, not all of them, exceeded the one-second threshold.
And if you look at why, there's another test checking tick volume. You can see that shortly after the open, as is often the case, there was a spike in record volume, which caused the latency spike. If you want to see the actual test,
you scroll down and see the results here. This is the test that failed, and if you scroll to the very end, you get a lot of detail on the mounts that failed. Then you can go back to the dashboard itself.
If you want a summary of the issue, you can see that the MEMDB loader latency exceeded the 1,000-millisecond threshold for thread zero on mount 6, mount 7 and mount 9, so three mounts.
That's how you'd find and quickly diagnose a data quality issue. When this test failed, it would have sent an alert by email as well as Slack.
Anybody monitoring that email inbox or Slack channel would have gotten the notification. Since I'm running a little short on time, the last thing I want to show you is the AI integration within TrueTick Studio. TrueTick Studio is the authoring environment, and one of the challenges of onboarding data is writing these notebooks.
The data is often very rich, and the person writing the data quality notebook may not be the expert on that data. So it's very useful, first, to catalog the data through the TrueTick data catalog, so that both the agent and you can inspect the available databases and schemas.
For example,
I asked an agent, and not just any agent but Claude, which is included, to help me create a data quality test plan for an options database. As an experiment, I provided it with a set of best practices for pricing options that I found on LinkedIn.
These were rules such as: remove contracts where implied volatility cannot be solved because Black-Scholes fails to converge; remove duplicate quotes; filter abnormal breaks between delta, gamma and theta. These are the invariants, or requirements, that I provided to Claude as input to the test cases I wanted Claude to design, and ultimately write, for me. With a little prompting, and using the MCP server, skills and data catalog, for example this list databases call, Claude found that there's a US options end-of-day database in OneTick.
And now it knows the schema.
It discovered the actual fields. Then it found this stats table, which has information about calls, puts and expirations by contract. It found quotes, and so on. It discovered all the data it needed via the TrueTick catalog and the TrueTick catalog MCP server.
It did some more investigation, determined that it needed rates, and found the rates it needed. Really, it took just a few minutes. The end-to-end process was maybe an hour or two, because there was some validation I wanted to do. It came up with a data quality test plan defining three end-of-day notebooks: one for equity options, one for ETFs, and one for index futures, because, for example, pricing futures uses a different pricing calculation than equities; futures have cost of carry to account for, and so on. It identified all the databases in scope for testing, either as a source of input data or as the data being tested itself, and the symbols that should be used, and then it proposed all the tests that should be written to meet those requirements.
TrueTick supports dependencies between tests. For example, you wouldn't want to test that a partition exists in a particular archive if the load didn't happen. Only if the load happened would you check that the tick types are available, and only then check that the schema is valid, and so on. This prevents cascading errors.
Here's the rest of the plan for all the tests. You can then review the plan, tell Claude it's okay, and have it write your notebooks.
That's an example of the power and utility of integrated AI assistance and the data catalog, which is highly integrated into TrueTick.
I think I've used up my time, so I want to allow some time for questions.
I'll close by asking: where do you stand?
If any of this sounds familiar, like "we don't do automated testing," "we manually investigate breaks only when they happen," or "we frequently learn about data quality issues from our users," then I strongly encourage you to come talk to us. I'd love to help you apply TrueTick to make sure your data is in fact trustworthy.
Again, here's the link to learn more about TrueTick: kx.com/product/truetick. We'd love for you to join our private preview. If you do, we'll work with you to stand up your own TrueTick stack and assign an engineer to implement TrueTick with your data, at least one database, so you can experience firsthand how TrueTick can ensure your data is trustworthy and AI-ready.
Alright. Any questions?
Q&A
Justin Fry: Mick, thank you so much for that excellent presentation and a great demo. We've got a few questions. The head of market data at a multi-strategy fund asks: "We load files from several vendors every night, and not all of the data is in OneTick. Can TrueTick test data that isn't coming from OneTick?"
Mick Hittesdorf: Yes, it can. TrueTick today is designed to work with OneTick, and there's essentially a OneTick database server integrated into TrueTick. OneTick can read, among other file formats, Parquet, both Parquet in the OneTick format and arbitrary Parquet.
Also, if you're using S3 as your data lake, as many firms are, TrueTick can read data directly from S3. So there are definitely options if you're not an existing OneTick customer and your data is in another format. Especially if it's in Parquet, we can definitely accommodate you.
Justin Fry: Thank you, Mick. Someone else asks: "Monitoring tools bury us in false alerts. How do we avoid that with TrueTick?"
Mick Hittesdorf: Great question. I touched on it briefly earlier. TrueTick is calendar-aware. That's one of the advantages of OneTick Cloud's rich data and reference data; we are a market data vendor.
We've invested heavily in calendars, for example, and we know all about the different sessions: the pre-open, the open, after-hours. There are even markets in Asia that close for an hour during the day for lunch.
We know all about those market idiosyncrasies. So, for example, say you schedule a job to run Monday through Friday. If a Monday, or Good Friday, is an exchange holiday, or it's a half day, TrueTick knows from the calendar and won't run the job. That's a big source of false positives.
Another capability that minimizes false positives: even if a test fails, you can use a pytest decorator to convert a failing test into what's called a warning result, and warnings don't trigger alerts. That's very important for tests that are statistical or probabilistic in nature, or where you're aware of an upstream issue you haven't been able to solve yet. Maybe the test fails once or twice, you know the source of the problem, but you haven't yet been able to coordinate with the upstream data owner to fix it. You just add one line to the test in Python.
The test result still gets logged, but as a warning, so you can keep tracking it, and alerts are turned off for that test. Those are two examples of how TrueTick intelligently avoids a lot of false positives.
Justin Fry: Thank you, Mick. A quant researcher at a systematic equities firm says: "I already run my own sanity-check queries before I backtest. Why would I change? Why would I use TrueTick?"
Mick Hittesdorf: First, I want to applaud you for doing that. I think you're in the minority of firms that do those sorts of sanity checks.
One thing I'd love to talk to you about: are other people within your firm benefiting from what you're doing? Do you have a way of managing your tests and tracking them systematically? Do you have sophisticated tools for diagnosing failures? When your tests fail, can you alert people who can actually do something about it?
And who helps you write those tests?
Do you have good AI assistance, and the ability to interrogate the data you're relying on?
You may have figured out ad hoc solutions to some of those, but TrueTick addresses those requirements comprehensively, and at the same time doesn't restrict you from writing your own tests. One of the beauties of TrueTick is that, at the end of the day, it's a Jupyter notebook that lets you write standard Python and SQL. It's very common for end users who do the kind of sanity checks you're doing to know that if a certain query passes, the data is good for their purposes, and if it fails, it isn't. TrueTick lets your query be systematically incorporated into a job, and lets you as the end user look at a dashboard, see that your test is being run in an automated fashion, and track the results.
So I'd love to talk to you about what you're doing. I think we can build on the work you may already be doing.
Justin Fry: Thank you, Mick. A chief data officer at an asset manager asks: "How do I show the business that data quality is actually improving over time? And is there a timeline I should expect, so I can set expectations about how things will change?"
Mick Hittesdorf: I think I showed that during the demo. The data quality dashboard lets you do trend analysis not only at the data estate level but for a specific database. Let's choose a common one, US Comp, our US equity database.
You can choose a particular database and time range. I'll turn this filter off, since it's showing only the failed tests. In this case, you can see the data quality is pretty stable and has been good for the last ten days.
If you look at the last thirty days, you can see there were some issues here. Even this is very valuable: you can show a stakeholder that you're running over 1,500 tests a day, both T+1 and real-time, and that your data quality scores are consistent.
And if they ask, "What about the seventh? We had some issues," you can drill in and say, "Yes, we had some issues."
"This was the cause, and this is what we did to remediate it." That level of transparency lets you have an honest, data-driven, quantifiable and visual conversation with your stakeholders. For new databases, you'll often see a trend line that starts below the KPI, improves over time, and at some point stabilizes, hopefully at a level where the data is consistently high quality and meeting the KPI.
Justin Fry: Thank you, Mick. We're down to the last few questions. A head of AI at a bank asks: "How would we know our data is AI-ready?"
Mick Hittesdorf: Wow, the $64,000 question. I think that's on everyone's mind.
What I'm hearing in the industry is that we've gone beyond experimenting with AI and trying to understand what LLMs can and can't do.
Many pilots have been conducted at capital markets firms and in other industries, and what we've observed is that what prevents pilots from going into production is a lack of high-quality, AI-ready data.
That raises the question: what is AI-ready data? It's data that addresses what I call the failure modes of AI.
When I say AI, I mean both traditional machine learning, which capital markets participants use quite extensively,
especially for training models, and LLMs. They have different failure modes. We'll be releasing a white paper on AI readiness and how to quantify it, but it builds on the TrueTick data quality score foundation: a data quality score from tests for those failure modes. For machine learning models, that's things like look-ahead bias and survivorship bias. For LLMs, it's contradictory context and temporal confusion, where something that's right in one time frame isn't right in another.
For example, symbols change over time. There's a corporate action, a split or a merger, and that can be very confusing to an LLM that thought it was talking about one thing.
Say Walmart is trading at a certain price level, then undergoes a three-for-one split, and now it's at a different price. If the LLM doesn't know about the corporate action and the date it happened,
it may think the stock just plunged. Looking at the price history alone, it looks like the stock crashed, but it was just a split. So there are failure modes unique to machine learning and AI, and you can design an AI readiness test suite that looks for the conditions that lead to them. That's how you translate a data quality score into an AI readiness score.
Once your AI readiness score reaches a certain threshold, whether that's 95%, 99% or whatever, you have an objective measure of whether your data is AI-ready.
Justin Fry: Quick question here from a market data operations manager at a broker-dealer who joined a bit late: "What does it take to get this running? Do I need a project team, or could one engineer implement TrueTick in a week? How should I plan for this?"
Mick Hittesdorf: We run TrueTick in SaaS mode as part of OneTick Cloud, and our hosted solution is typically the fastest way to start using TrueTick. It's hosted by us, but we stand up a dedicated stack for your organization.
We take data security very seriously. We have an automated process that can deploy a TrueTick stack within hours or days, depending on whether you have any unique requirements; typically within hours. The only requirement is that our stack can reach your data. If your data is in AWS, that's just a matter of working with your IT security team to set up cross-account resource sharing.
Again, we're very serious about security, and we'll work with your IT security team to ensure the connection to your data is secure. If you're running in a different cloud, that's possible too: we have a customer running in Azure, and we use a Direct Connect to connect to and query their OneTick instance.
And if you want to run TrueTick within your own cloud account, because data sovereignty and control are paramount to you,
TrueTick will shortly, likely next quarter, be available for what we call BYOC, bring-your-own-cloud deployment. SaaS mode is the quickest way to get started today.
That's the option we're offering to private preview customers, and we'll offer a BYOC option in the near future.
Justin Fry: The head of market data at an asset manager asks: "Can I use this to hold my data vendors to their SLAs and show them evidence when they miss?"
Mick Hittesdorf: Absolutely, and I'd encourage you to do so. As I mentioned earlier, customers rarely do this. When you buy data from vendors, you should be asking how they validate their data and what methodology they use. I'd go so far as to ask: do you score your data? Do you have an objective measure of data quality, and how is that score computed?
One of our goals with TrueTick is to raise awareness of data quality and, ideally, to standardize how data quality is measured across vendors. You should definitely challenge your vendors on data quality and not make any assumptions about it. Please reach out; we'd be happy to talk with you about TrueTick and how it can help with your data quality challenges.
Justin Fry: Thank you, Mick, and thanks to the audience. We look forward to seeing you at the next event. We'll send out a recording of this. As always, to learn more about KX, OneTick and TrueTick, please visit kx.com/product/truetick.
