Insights

JSE's New Algo Trading Rules Create an AI Governance Obligation No SA Bank Has Mapped

Terry Yodaiken ·

Forty-five per cent of South African banks intend to spend more than ZAR 30 million on AI this year. Two years ago most of them spent under ZAR 1 million. Those two figures come from the joint SARB and FSCA study of AI in the South African financial sector, and the multiple between them is roughly thirty. Governance headcount has not moved by anything like the same factor.

Into that gap, in May, the JSE dropped a new set of algorithmic trading rules.

Research and drafting for this article were AI-assisted. Every figure is linked inline to the source that published it. Editorial responsibility is mine.

A clean white decision flowchart with yes and no branches on the left, divided by a vertical gold line from a dense, glowing cyan neural network on the right.
Produced by Gemini-Nano-Banana

Rules Written for Deterministic Code

The exchange lifted its algo trading and direct-market-access controls out of technical directives and into the formal rulebook, which gives it stronger enforcement powers and puts explicit senior management accountability on algorithm design, testing and monitoring. Brokers offering direct market access carry responsibility for their clients’ algo activity, including where a third-party vendor built the system. Pre-trade controls are mandatory.

Sensible, overdue, and closely modelled on what MiFID II established in Europe more than a decade ago.

The difficulty is what the rules assume. Algorithmic trading is defined as computer programs or AI buying and selling automatically according to predefined rules, and every governance requirement that follows is built on that definition. You design the algorithm. You test it. You document how it works. A senior manager signs it off. The exchange knows what it is supervising because the thing being supervised holds still.

Machine learning does not hold still.

A gradient-boosted execution model trained on six months of order flow follows no predefined rules in any sense a lawyer would recognise. It infers statistical relationships and applies them probabilistically. Retrain it, as production models routinely are, and the parameters change. Tuesday’s model and Thursday’s model are not the same artefact, and neither one was designed in the way the rulebook imagines designing something.

What ESMA Said That the JSE Did Not

Europe met this question head-on in February. ESMA’s supervisory briefing confirms that AI and machine learning strategies fall inside the MiFID II algorithmic trading definition wherever they autonomously determine one or more parameters of an order. It goes further, telling firms using machine learning in order management or execution that governance, testing and audit-trail documentation must meet the same expectations, and singling out reinforcement learning, deep learning, neural networks and generative AI as raising distinct risks. RTS 6 annual self-assessment sits underneath all of it.

The JSE’s rules contain no equivalent passage. They import the accountability architecture without the acknowledgement that the thing being held accountable has changed shape.

That is not a criticism of the exchange so much as an observation about timing. The changes were first proposed in March 2024, and the trading estate they were written for is not the one they now govern.

Three Rulebooks, One Desk

A CIB head of trading in Johannesburg now carries three concurrent obligations over the same models, issued by different bodies on different theories, with nothing connecting them.

The JSE rulebook names senior management as accountable for design, testing and monitoring, holds brokers responsible for third-party algorithms, and requires pre-trade controls. It offers nothing on continuous learning, retraining cadence or probabilistic output.

The Prudential Authority’s D12-2025 credit risk roadmap requires banks on the internal ratings-based approach to obtain prior written approval before any material model change. A retraining cycle is a model change. The SARB and FSCA study conceded that most local independent validation functions have not built machine learning validation capability in-house, and the PA has given no indication it will relax validation rigour to accommodate that. For a desk running ML-driven counterparty credit, potential future exposure and CVA models, every retraining event becomes a governance event requiring an approval the bank may not have the internal capacity to produce.

The AI Joint Standards still to come from the FSCA and the PA will add a third layer. The joint study signalled sector-wide guidance on ethical, fair and responsible AI, with a discussion paper preceding formal consultation. Timing is unconfirmed. The direction is not: principle-based requirements sitting alongside, and quite possibly across, the rule-based obligations already in force.

No South African bank has published a framework mapping the three onto one control set. Market risk owns the exchange rules. Model risk owns the PA. Compliance is waiting for the standards. Nobody owns the joins.

The Surveillance Blind Spot

Exchange surveillance, at the JSE as everywhere else, is built to detect known manipulation signatures. Spoofing, layering, front-running, wash trades, marking the close. These are pattern-recognition problems with defined shapes, and platforms like Nasdaq’s SMARTS, deployed across more than fifty exchanges and 190 banks, are good at finding them.

Learned models fail in ways that do not match those shapes.

Take three South African banks deploying execution optimisation models trained on overlapping JSE order-flow data. Each learns similar statistical relationships, independently, with no intent of any kind. The correlated behaviour of three independently-learning systems could produce a price impact pattern indistinguishable, from the surveillance desk’s seat, from coordinated manipulation.

So who answers for it? Surveillance sees a pattern and flags possible manipulation. Model risk sees three models operating inside validated parameters. The JSE’s rules require a clearly accountable party responsible for managing the risk, and each firm can demonstrate its own algorithm was designed, tested, monitored and signed off exactly as required.

The exposure was created between the firms, and every regime here locates risk inside one.

Validation Capacity Is the Binding Constraint

Basel III, as the PA operationalises it, requires model validation by a function independent of model development. For a traditional scorecard that worked well enough: the validation team could read the coefficients, reproduce the arithmetic and argue with the methodology.

For gradient boosting, neural networks and reinforcement learning, the same team needs a different discipline entirely. A published guide to the PA’s validation expectations sets out what that involves in practice: SHAP-based explainability analysis, hyperparameter assessment, population drift detection where a stability index above 0.25 signals a significant shift, and discriminatory power evaluation where a Gini coefficient of 0.40 is treated as the floor for acceptable performance.

Very few second-line teams in the local market can do all four today. The PA’s own guidance acknowledges the shortfall, and the D12-2025 roadmap describes phased implementation for banks to assess model gaps without extending any deadline. The capability gap has to close while deployment accelerates, and the exchange has now stacked algo-specific obligations on top.

ENSafrica named the underlying problem when the joint study came out: institutions lack comprehensive governance structures, including data governance, model risk management and board-level oversight. Set that beside the constraint ranking in the same data, where 82% of institutions name data protection as their leading AI constraint and 48% name market conduct legislation, and the misalignment is plain. The constraint firms are worrying about is not the one the new rules create.

Three Questions Worth Answering Before Somebody Asks Them

The gap between the rules being in force and the AI standards arriving is the only period in which any of this can be designed rather than retrofitted.

Which systems are actually in scope? Not just execution algorithms. Pricing models, risk models, and anything that autonomously determines a parameter of a trading decision. Most desks will find the list longer than the one market risk is currently maintaining.

Does a retraining cycle require fresh sign-off? Under the exchange rules that is genuinely unclear, and under D12-2025 it looks like a material model change. Write the internal policy now, because the alternative is having one written for you during an examination.

And who owns the interaction? Between the three regimes, and between your models and everyone else’s. Right now, at most South African banks, the honest answer is nobody, and that is the answer a supervisor will eventually be given in writing.

The rules assume the risk sits inside the firm. Increasingly it does not.

Regulatory position as at June 2026. The FSCA and PA AI Joint Standards had not been published at that date and no consultation timetable had been confirmed.

Where These Figures Come From

AI GovernanceSouth AfricaAlgorithmic Trading