When Algorithms Run the Market

Kunal Kumar

Over sixty percent of trading activity in modern markets is now driven by algorithmic systems. That's the number in our WelthWest piece on AI in trading. Sixty percent. Humans are the minority in the room.

Shreya

And the room is mostly servers talking to servers.

Kunal Kumar

Right. And my, my thesis is this. Growth in fintech doesn't get measured by how fast the model is, or how much volume it pushes. It gets measured by the guardrails nobody sees. Whether they hold.

Shreya

Okay. Prove it.

Kunal Kumar

August 1, 2012. Knight Capital. So, this part is from the SEC's own findings. In 2005, Knight moved a section of code in its automated order router, and that broke an old function called Power Peg. The function wasn't meant to be used. They just left it in there.

Shreya

Left it sitting.

Kunal Kumar

For seven years. Then late July 2012, they deploy new code for the NYSE's Retail Liquidity Program. And per the case study I read, one of their eight servers never got the update. Market opens, orders hit that one server and the old function wakes up.

Shreya

And the router can't tell when an order has been filled.

Kunal Kumar

That's it. So it just keeps firing. The SEC says in the first 45 minutes, it sent more than 4 million orders into the market, trying to fill just 212 customer orders.

Shreya

Two hundred twelve. Against four million.

Kunal Kumar

And the SEC says Knight traded more than 397 million shares, and acquired several billion dollars in unwanted positions. Loss, more than $460 million. Other accounts say about $440 million pre tax, so, um, I'll flag that range.

Shreya

From a balance sheet perspective, that is the company. Wikipedia, citing Associated Press reporting, says that by the next day, seventy five percent of Knight's equity value had been erased.

Kunal Kumar

And here's the part that gets me. The SEC order says an internal system sent 97 automated emails about the router before the market even opened.

Shreya

Ninety seven.

Kunal Kumar

Nobody acted on them.

Shreya

Then that's not only a technology failure. It's a listening failure. But let me push on your thesis. The industry pitch today is that the models are smarter. Language processing on earnings calls, sentiment feeds scanning thousands of sources in real time. The implication is that operational risk got engineered out.

Kunal Kumar

I'd say it got, uh, it got faster. Not smaller.

Shreya

That's my view too, and here's the mechanism. Smarter prediction and better containment are different skills. A sentiment model improves what you decide to trade. It does nothing for a stale server. And at microsecond execution, one latent error is copied across the market before a human risk officer finishes reading the first alert.

Kunal Kumar

Fair. Though our own source piece does list the limits. Struggles with unpredictable events, depends on data quality, can fail when the market regime changes.

Shreya

The disclaimer is always at the bottom.

Shreya

So let's talk about who's building on this. Your source names a few platforms. Alpaca, Numerai. They're not the same bet at all.

Kunal Kumar

No. Alpaca is API based trading automation for developers. You bring the code, they give you the pipes. Numerai is an AI driven hedge fund using crowd sourced models.

Shreya

And that second structure is interesting to me. If thousands of independent data scientists each build a model, then one model collapsing doesn't take the fund down. That's diversification applied to the algorithm itself, not just the assets.

Kunal Kumar

Spreading the model variance around. Yeah. Whereas with the developer pipe, the risk sits with whoever wrote the code. Which, um, brings me to something from our own scaling.

Shreya

Go on.

Kunal Kumar

We had an automated rebalancing routine running during a volatility spike. A big one. And the backtested math said, keep going, the routine is behaving as designed. But the exposure on the balance sheet, it, it didn't look like the math. So I wanted to halt it by hand.

Shreya

And the engineers?

Kunal Kumar

Pushed back. Hard. Their argument was, the model has been tested on years of data, and you're overriding it on a feeling. And they weren't wrong that it was partly a feeling. But the feeling was attached to a number I could actually see.

Shreya

That's the right instinct, though I'd separate two things. Overriding because you're nervous is one thing. Overriding because live exposure diverged from the assumptions is a control working as intended.

Kunal Kumar

Yes. Okay, yes, that's cleaner than how I said it.

Shreya

Which loops back to your opening. The backtest only knows what already happened. In a real regime break, the history has no precedent, so the model is confidently extrapolating from a world that's gone.

Kunal Kumar

And Knight is the extreme version. Forty five minutes, no kill switch fast enough. The case study I read lists the lessons pretty plainly. Human oversight, override capability, escalation that someone actually follows.

Shreya

So where does that leave AI in wealth management?

Kunal Kumar

As an execution partner. It does the rebalancing, scans the data, never gets tired. But a human owns the kill switch, and the switch has to be rigid. Not a suggestion in a dashboard.

Shreya

The future isn't manual versus AI, as your piece says. It's whether the human in the loop can actually stop the loop.

Kunal Kumar

And whether anyone reads the 97 emails.