AI Scientists Simulate the Jamaica Stock Exchange With Generative AI

A research team at Maestro AI Labs has built HS1, an AI model that simulates the Jamaica Stock Exchange. The name is short for Harbour Street, the Kingston home of the exchange it models. The team was mostly financial professionals and market experts rather than a room of AI engineers, though it had those too. HS1 learns from more than a decade of historical trading data, takes economic-model inputs alongside that price history, and uses generative AI to produce forward paths for market behaviour seven to fourteen days out. It covers every listed entity on the exchange, not a handful of liquid blue chips, and it is assembled from a mix of open-source agents and financial machine learning models the lab built itself.

What will matter in five years is not the forecasting but the second half of the announcement: Maestro AI Labs says the project came out of a need to support financial literacy and informed investment decisions, and that HS1 is being built into a game engine, free to use, so that every Caribbean citizen can learn investment and finance in a zero-risk environment. A regional lab modelling a regional exchange and then turning it into a public teaching tool is a first for Caribbean capital markets. What it becomes depends on what the lab ships and what evidence comes with it.

TLDR

• Maestro AI Labs has trained HS1, a generative model, on the Jamaica Stock Exchange, using over ten years of trading data plus economic-model inputs to simulate market behaviour seven to fourteen days forward across all listed entities.
• By the lab's own numbers, HS1 is accurate to about three days, decaying by day eight, a fall-off that is hard to avoid where many counters barely trade.
• Open-source AI agents handle the research and workflow while custom machine learning models do the simulating, weighing price history, economic conditions, news and media, and even physics equations and other less obvious factors. The team was mostly financial professionals and market experts, not only AI engineers.
• The lab frames the purpose as financial literacy, and is building HS1 into a game engine, free to use and aimed especially at young people, so Caribbean citizens can practise investing without risking money and put it into things they believe in.
• The JSE is an unusually hard market to model, with thin trading, small free floats and long stretches where a listed counter does not trade.
• CAIA's read: the evidence matters more than the forecast. Published backtests against naive baselines, a stated holdout period, and honest error rates are what would turn this from an announcement into a teaching tool the region can trust.

What the Model Does

The inputs explain the design. The first is more than a decade of Jamaica Stock Exchange trading history: prices, volumes and the record of when each listed security changed hands. The second is a set of economic-model inputs, the macroeconomic conditions that sit underneath any equity market and that a price series on its own cannot see. A model fed only prices learns the shape of past moves; add the economic state those prices moved through and it has a chance of telling a shift in interest-rate expectations apart from one large investor rebalancing a portfolio.

The output is a simulation rather than a single number. Generative models produce sampled paths, and running many of them gives you a distribution of where a price could plausibly sit in seven to fourteen days rather than a point estimate that is wrong by construction. On the lab's own reckoning, HS1 stays accurate to roughly three days out and falls away by day eight, close to the degradation you would expect on a market this thin. That distinction gets lost easily in coverage of work like this, and it decides whether a reader comes away understanding how uncertain a market is or believing a number.

The lab took the assembly route with HS1: open-source AI agents run the research and the workflow, handling collection, cleaning, feature preparation and orchestration, while the forecasting itself sits in purpose-built machine learning models. Those models are what simulate the market, weighing price history and economic conditions alongside news and media, even physics equations and other factors that would not obviously bear on a share price. In a small market that split makes sense. Most of the engineering time on a project like this disappears into the agent layer, where open-source tooling has become good enough to lean on. The models are where a Caribbean market needed its own answer, because nothing trained on New York or London order books has seen anything that behaves like Kingston's.

Why the JSE Is a Hard Test Case

The Jamaica Stock Exchange opened for trading in February 1969 and now runs a Main Market, a Junior Market created in 2009 to bring smaller companies to public capital, a USD market and a bond market, together carrying well over a hundred listed securities. Bloomberg named it the world's best-performing stock exchange in early 2019, and it migrated its trading and surveillance onto Nasdaq technology later that year. By the standards of what a modelling team needs, it is a well-run exchange with a data history worth training on.

It is also a frontier market, and frontier market microstructure breaks assumptions that developed-market models are built on. Many counters trade rarely. A closing price on a security that last changed hands four days ago is an artefact rather than a current valuation, and a model trained naively on daily closes will read that stillness as low volatility, learning that Jamaican equities are far calmer than they are. Free floats are small enough that a single institutional block can move a price several percent without any new information entering the market. Volumes behind a quote are a fraction of what the same model would see on a developed exchange, so the statistical power available per counter is thinnest where the modelling problem is hardest.

All of that makes the JSE a serious research target. A model that handles non-trading days honestly, tells a price move apart from a liquidity event, and is candid about its own uncertainty is solving problems most published equity forecasting has been free to ignore. It also explains why Caribbean frontier markets have attracted so little AI-driven analytics. The data is harder to work with, the market is smaller, and getting it right pays a fraction of what the same effort earns in a larger one. That is the argument for a regional lab doing the work, and for a regional association paying attention when one does.

The Literacy Argument

The lab's stated motivation is financial literacy, and on that ground the design choices make more sense than they would for a trading product. The aim it describes is informed conviction: enough understanding that a first-time investor can put money into things they believe in. Jamaica has spent a decade widening retail participation in its equity market, with the Junior Market bringing a generation of smaller companies and first-time shareholders into public capital. That grew the number of owners faster than it grew their understanding of what they own. Most first-time shareholders learn what a market does by watching their own money move, which is an expensive and distorting way to learn anything.

A simulator changes what a beginner can practise. Someone can hold a position through a simulated fortnight, watch a thinly traded counter sit motionless for four days, and see what a block trade does to a small float, without paying school fees to the market for the lesson. Done well, it teaches two things a first-time investor is rarely told: that short-horizon prices are mostly unpredictable, and that the width of the range of outcomes carries more information than its midpoint. The lab is building HS1 into exactly that: a game engine, free to use and aimed especially at young people, where a beginner can build the skill before bringing real money to a real market.

The same design carries a risk. A simulator built to teach humility about forecasting can just as easily be read by a beginner as a machine that knows what happens next, and which reading wins comes down to interface and framing more than model quality. It turns on whether the tool shows a fan of outcomes or a single line, whether it puts its historical error next to its forecast, and whether it says plainly, right where someone would act on it, that this is a teaching environment and not advice.

What the Lab Still Has to Show

A free teaching tool raises the bar on evidence rather than lowering it, because the people using it are beginners who cannot check the model themselves. What makes the work trustworthy is its track record: the walk-forward backtest, the exact boundary between the data the model trained on and the periods held back to test it, and the error rates on data it never saw. Without those, a beginner has no way to know whether the fortnight they just watched was skill or noise.

There is also the question of what the tool teaches when it is wrong. A model that shows a confident path and turns out to be off by day eight can teach the wrong lesson as easily as the right one, so the game should carry its historical error alongside every forecast and say plainly where its accuracy runs out. The lab has been candid about the three-day figure; carrying that candour into the interface, where a first-time investor sees it, is what would make the tool worth learning from.

The Regulatory Line

Jamaica's Financial Services Commission licenses investment advice, and the boundary between an educational simulator and an unlicensed advisory service is a matter of presentation as much as intent. Show a distribution of simulated outcomes for a security, label it clearly as a training environment, and the tool sits on the education side of that line. Wrap the same model in an interface that ranks tomorrow's best counters and it has crossed to the other. Any team releasing this publicly should expect that question early, and should have answered it inside the product before anyone asks from outside.

There is a second-order issue specific to thin markets, and it is worth naming before it happens. If a public forecasting tool becomes popular enough on a market where a few million dollars can move a price, the forecast starts to influence the thing it forecasts. That reflexivity is negligible on a deep exchange but a live design constraint on the JSE, and the mitigations are unglamorous: publish uncertainty prominently, resist ranked buy-side outputs, and watch whether release-day volume starts to track the model's own forecasts.

What CAIA Will Be Watching

Four things will show whether this becomes a teaching tool the region can trust or a well-received announcement. Start with out-of-sample results against naive baselines. The one that matters most is the last-traded-price benchmark, hard to beat over a seven to fourteen day horizon on an illiquid counter; a model that beats a random walk on Jamaican equities, on data it has never seen, would be a result worth publishing on its own terms.

Calibration comes next. A generative simulator only proves useful if it gets the shape of uncertainty right, so the lab should report whether outcomes fall inside its stated intervals at the rate it claims. Then there is what ships alongside the game: the backtests and the error rates, or a polished interface wrapped around a model nobody outside the lab can check. Last is what happens with the exchange and the regulator. A simulator trained on a decade of JSE data and handed to the public as a learning tool works better with the exchange than around it, and the JSE has run investor education programmes long enough to be the obvious partner for a classroom-facing version.

None of this diminishes the achievement. Frontier markets are under-modelled because they are hard and small, and the region has spent years importing analytics trained on markets that behave nothing like its own. A regional lab training on regional market data, to teach regional citizens how their own markets work, is the direction CAIA has argued for across every sector it covers. The work exists; the evidence now needs to catch up to it, and if the lab publishes the backtests behind the game, it will.

Related Reading Across the Caribbean AI Network

• Jamaica's AI Moment: from the call centre floor to the cutting edge
• AI Literacy in the Caribbean: what every citizen needs to know by 2027
• StarApple AI, the Caribbean's first AI company
• Caribbean AI Risk Management Council, on AI governance and risk across the region