ChronexaINVESTMENT RESEARCH & PORTFOLIO DATA AUTOMATION
Nobody serious is buying a system that picks stocks. What actually costs firms their week is consolidating positions across accounts they do not custody, running the same screen by hand every quarter, and assembling review packs. That work is systematisable. The other thing is not.
Plaid, Yodlee, custodian feeds and market data, in your own accounts.Built so every output can be reproduced later — which regulation increasingly requires.
“Show one household’s true position across everything, including what we do not custody.”
What goes in
What is computed
What a person decides
The adviser reads a household position that is actually complete, including the held-away accounts that usually make the picture wrong.
ON THE RECORDEvery source, the retrieval timestamp, and each reconciliation decision — so a number in a client review can be traced back months later.
01 / THE CONSTRAINT THAT SHAPES EVERYTHING
In most industries overclaiming about AI is embarrassing. In this one it is an enforcement matter, and it does not require anything to go wrong with the portfolio — only for the description to outrun what the records support.
That is not a disclaimer at the bottom of this page. It is the reason the system is built to reproduce every run.
In March 2024 it brought actions against two investment advisers for statements about AI and machine-learning capabilities they did not actually have. The enforcement risk here is not hypothetical and it does not require anything to go wrong with the portfolio.
The Marketing Rule governs advertisements including AI-generated content and any performance figures attached to them, and it prohibits untrue or unsubstantiated statements and cherry-picked results. A capability you cannot evidence is a marketing problem before it is a technology problem.
The compliance-rule requirement is for written policies reasonably designed to prevent violations. If AI is being used materially in your process, policies that are silent on it are not reasonably designed — and silence is the default state at most firms right now.
Books-and-records obligations mean the material behind a performance figure has to be retained. That has a direct architectural consequence: every run has to be reproducible, with its data vintage, its criteria and its output kept — which is why that appears on every tab of the demo above.
02 / WHAT WE USE AND WHAT WE REFUSE
Most agencies list every model they have heard of. It is more useful to say where each one genuinely earns its place, and where reaching for it is the tell that someone is selling.
Genuinely strong on the tabular, feature-rich problems this industry actually has: ranking a universe against stated criteria, classifying transactions, scoring data quality, flagging which filings deserve a human read. Explainable enough that you can say why something ranked where it did.
We use theseReading far more source material than a team can, and pulling out what matters with the passage attached. This is where AI most reliably adds hours back to a research process, and it is the least glamorous thing on this list.
We use theseDefensible for specific, well-posed problems — anomaly detection in flows, filling gaps in data, some volatility work. We have used them where the problem genuinely had that shape, and we will tell you when yours does not.
SometimesOut-of-sample performance on raw price prediction does not survive contact with reality, and marketing it is the exact conduct the SEC has charged firms over. We will build your screening, risk and research infrastructure. We will not build you a claim you cannot substantiate.
We decline03 / THESE ARE NOT THE SAME BUSINESS
Vendors sell all of these the same product. The honest position is that the problem changes completely depending on which of these you are — so find yourself here before reading anything else on this page.
You cannot advise on what you cannot see, and the client’s real picture is spread across accounts you do not custody. Consolidation, drift monitoring and review preparation are where the time goes — not stock selection.
More filings, transcripts and factsheets exist than any team can read. The work is triage: getting the right twenty documents in front of an analyst with the relevant passages marked, consistently, on a schedule.
Identifiers that disagree across providers, corporate actions that break history, feeds that fail quietly. The unglamorous plumbing decides whether anything built on top is trustworthy.
Multiple custodians, private holdings that no feed covers, reporting that has to match how this family thinks. Off-the-shelf portfolio software tends to fit badly, which is precisely when a built system earns its cost.
This is product engineering rather than internal tooling — aggregation, categorisation and analytics that have to work for every customer, not just the ones your ops team can babysit.
Firms looking for a model that tells them what to buy are looking for something else, and probably something that does not work. We build the infrastructure around an investment process that already exists.
04 / THE DATA LAYER
A Plaid link takes an afternoon. Making the resulting numbers trustworthy takes the rest of the project, and that is the part nobody demos.
The same holding arrives under different identifiers from different providers. Corporate actions break historic series. Fund composition is not always current. Feeds fail in ways that look like a genuine change in position rather than an error. A consolidated total that quietly drops what it could not reconcile is worse than no total, because somebody will put it in front of a client — so ours reports what it could not reconcile, by name.
05 / THE ENGAGEMENT
Usually consolidation, because everything downstream depends on the positions being right. Getting that wrong makes every screen, every risk number and every client review built on it wrong in a way nobody notices for months.
RELEVANT WORK
We have built account aggregation and market-data workflows across US and Indian markets — Plaid and Yodlee for held-away accounts, exchange and fundamental data, fund composition for look-through — and used gradient-boosted models where the problem was genuinely a ranking or classification problem rather than a prediction one.
Our own co-founder is a chartered accountant and SEBI-registered investment adviser, which is a large part of why this page argues about substantiation rather than about model architecture. On a call we will go through what we built, what it did, and where the reconciliation work turned out to be much larger than expected.
How the integration side works06 / START WITH THE POSITIONS
Every firm has one — the consolidated total, the exposure figure, the screen output that somebody checks by hand before it goes anywhere. That is the workflow worth building first, and it is a better starting question than which model to use.
You will be talking to the people who would build it, not an account manager.info@chronexa.io
Rather write it down first?
A FEW GOOD QUESTIONS
Because in this industry the regulation is the constraint that shapes the build, and pretending otherwise produces systems that cannot be used. In March 2024 the SEC charged two investment advisers over statements about AI capabilities they did not have, the Marketing Rule treats AI claims and their performance figures as advertisements subject to substantiation, and books-and-records obligations mean the material behind a figure has to be retained. Those three facts together mean a research system has to be reproducible by design — same data vintage, same criteria, same output — rather than having an audit trail bolted on later. That is an architectural requirement, not a compliance afterthought.
Gradient-boosted trees genuinely do. The problems this industry has are mostly tabular and feature-rich — ranking a universe against stated criteria, classifying transactions, scoring data quality, deciding which filings deserve a human read — and that is exactly where XGBoost and LightGBM perform well while staying explainable enough that you can say why something ranked where it did. Sequence models like LSTM are defensible on specific, well-posed problems such as anomaly detection in flows or gap-filling in data. What does not hold up is raw price prediction, where out-of-sample performance is poor and marketing it is the conduct that gets firms charged. We use these techniques where the problem has that shape and say plainly when it does not.
Not the connection. Plaid and Yodlee links are the easy afternoon. The hard part is that the same holding arrives under different identifiers from different providers, corporate actions silently break historic series, fund look-through requires composition data that is not always current, and feeds fail in ways that look like a real change in position rather than an error. A consolidated number that quietly drops what it could not reconcile is worse than no number, because someone will put it in front of a client. Most of the engineering in this work is reconciliation and failure handling, and it is the part that decides whether anything built on top can be trusted.
Usually not, and we would treat it as a warning sign if that were the first suggestion. Established portfolio and reporting platforms do a great deal well, and the gaps tend to be specific: held-away accounts that never make it in, a screen your process needs that the platform cannot express, look-through exposure the reporting does not compute, review packs assembled by hand every quarter. We build into those gaps and connect what you already run. Family offices with genuinely bespoke structures are the most common exception, because off-the-shelf software fits them worst.
It changes the data layer rather than the architecture. We have built across Indian and US market data — NSE and BSE alongside NYSE and NASDAQ — and the differences that matter in practice are corporate-action handling, identifier conventions, settlement and how fund composition data is published. The regulatory framing on this page is US-specific because most of the buyers reading it are; if your obligations sit with a different regulator then the specific rules differ, but the underlying design requirement does not. Reproducible runs and evidenced outputs are good practice everywhere and mandatory in most places.
It runs in your accounts, under your credentials, with your data-protection obligations designed in rather than assumed. That matters particularly for advisers because the amended Regulation S-P now requires written incident-response programmes and customer notification within thirty days of a qualifying incident, with compliance phased in across 2025 and 2026 depending on size. Practically, that means logging, retention and breach detection are part of the build rather than an operational afterthought, and we document what we built so your compliance people can review it without needing us in the room.
It starts with one workflow written down properly — usually consolidation, because everything else depends on the positions being right. That produces a scope, a definition of what counts as working, and a fixed price before any build starts. We would rather deliver one workflow that your team actually uses every week than a platform that impresses in a demo and is quietly abandoned by the second quarter, and pricing follows the number of data sources and workflows rather than a per-seat licence.