Michael Swift, Chief Artificial Intelligence & Quantitative Intelligence Officer at Buffint, describes the change in investment management as less about prediction and more about coverage. Models can read more instruments, more filings and more price history than any desk can review by hand, and the practical question is what a person does with that output.

The vocabulary is broad, so it helps to separate the tools. Machine learning finds patterns in tabular data such as prices, volumes and fundamentals. Natural language processing turns text into measurable fields. Optimisation routines turn a set of views into position sizes. Each answers a different question, and each fails in a different way.

Models Now Read Data That Analysts Could Not Cover

The first gain is breadth. A language model can scan earnings transcripts, regulatory filings and news headlines, then score tone, topic and change against the previous period. A classifier can sort thousands of instruments into groups that behave alike, which is a more flexible starting point than a sector label. Both tasks existed before, but they were limited by the hours an analyst could spend reading.

Swift is careful not to overstate what the technology does. "The model does not understand a company," he says. "It converts text and numbers into features, and it does that consistently across thousands of names. Consistency is the real contribution, because a human reader tires and a model does not."

The same logic reaches ordinary screening and alert tools. A screener filters on fields chosen in advance, whereas a learned model can weight many fields at once. The trade off is transparency. A filter that says price is above its fifty day average can be checked at a glance, while a score built from dozens of inputs needs documentation of what went into it.

A Signal Is Only as Good as Its Out of Sample Record

Quantitative work lives or dies on validation. Given enough variables and enough history, a model will find a pattern that fits the past perfectly and means nothing. This is overfitting, and it is the most common reason an impressive backtest disappoints in live trading.

The defences are well established. Data is split so that the model is trained on one period and tested on another it has never seen. Walk forward testing repeats that split through time, so the model is always judged on data that arrived after its training window. Transaction costs, slippage and the size of the order relative to available liquidity are then included, since a signal that disappears once costs are counted was never a signal.

Swift's team treats every impressive result with some suspicion. "A backtest is a claim about the past, and the first thing we ask of it is how many variations were tried before this one was kept," he explains. "If nobody can answer, the result should be treated as a hypothesis rather than a finding."

Automation Changes Execution Before It Changes Judgement

The most visible effect of automation is in how orders are placed. Algorithms can split a large order into smaller slices, work a limit price patiently and respond to changes in the order book faster than a person at a screen. Rules based triggers, such as a conditional order that activates when a level trades, are the simplest form of the same idea.

Speed brings its own risks. An automated rule repeats whatever it is told, including a mistake in a price field or a stale quote. Sensible set ups include maximum order sizes, price bands that reject an order far from the last trade, and a clear way to switch the process off. The control that matters most is often the least technical one, a defined limit on how much can be committed before a person looks again.

Governance Keeps a Model Accountable to a Person

Markets change regime, and a model trained in a calm period can behave badly in a disorderly one. Monitoring for drift, meaning a gradual gap between what the model expects and what the market delivers, is as important as the original build. Documented inputs, version records and a named owner for each model make it possible to explain a decision afterwards.

For Swift, accountability comes down to one question. "Who decided it should stop, and on what evidence?" he asks. "If that is written down before the first trade, the model is a tool. If it is not, the model is making the decision by default."

For an individual investor, the lesson scales down well. Decide the rule before using it, test it on data it has not seen, keep each position small enough to survive the rule failing, and log every change. Artificial intelligence widens what can be measured, and the discipline around it remains a human responsibility.

Michael Swift is Chief Artificial Intelligence & Quantitative Intelligence Officer at Buffint.com

Disclaimer: The content of this article is provided for general informational purposes only and should not be interpreted as personalised financial or trading advice. The author makes no representations or warranties regarding the accuracy, completeness, or timeliness of the information presented. Market dynamics are subject to frequent change, and past insights may not reflect current conditions. Readers should independently verify all facts and consult with a qualified financial adviser before making any investment decisions. The author and publisher accept no responsibility for any financial losses, decisions, or consequences resulting from reliance on this content. All actions taken based on this information are at your own risk.

COMTEX_494750984/2891/2026-10-23T03:13:01

You May Also Be Interested In

• Premarket movers: Crown Castle jumps 7.5% on SpaceX talks; Autoliv slides
• Autoliv shares fall 5% on third quarter earnings miss
• Archrock Announces Timing for Third Quarter 2026 Results