Unleash The Beast: What High-Frequency Trading Cost Us in Thirty-Three Days

There is a commit in our archive from April 2020 that reads, in full:

unleash the beast completely, force put, last write wins, TURN OFF LOCKING COMPLETELY SUPER HIGH FREQUENCY TRADING ACTIVATED!!!

Three exclamation marks. All caps. Thirty-three days later, the same person committed this:

lower frequency

This is the story of what happened in between, and it is the most expensive lesson in our history that did not involve losing a trade.

What "unleash the beast" actually meant

The build-up is visible over three days, and the escalation is unmistakable:

records locked debug message decorating HIGH FREQUENCY HIGH FREQUENCY 2 HIGH FREQUENCY 3 WARNING!!! unlock multi-threading WARNING!!! unlock multi-threading 2 TURN OFF LOCKING COMPLETELY SUPER HIGH FREQUENCY TRADING ACTIVATED!!!

Read what is being traded away. Locking is what stops two processes writing the same record at the same time. force put, last write wins means: when two updates collide, do not resolve the conflict — just let the last one overwrite the other and keep moving.

That is not a bug. It is a deliberate exchange: correctness guarantees, sold for speed. And the author knew exactly what he was doing, because he wrote WARNING!!! to himself — twice — in the commit messages immediately before.

The good

It worked, in the narrow sense that it did what it was built to do. The system trades faster when it is not waiting on locks. Anyone who tells you otherwise has not tried it.

And the instinct behind it was sound. In a market that moved as violently as crypto did in 2020, being early is worth real money. The reasoning — get there first, worry about tidiness later — is the same reasoning behind every low-latency trading operation on earth.

The bad — the retreat started the same day as the charge

Here is the detail that makes this story honest rather than triumphant. On the same day he turned locking off, he also cut the position size by roughly twenty to one.

Unleash the beast; then immediately bet much smaller. Those two instincts are in direct conflict, and both of them are in the archive, hours apart. Somewhere between writing the first commit and the third, the confidence had already started leaking.

Within a week, a throttle was back in the code. Within three weeks, this appeared:

disabling ... to test reduction of fees

That is the moment the bill arrived. Not a crash, not a bad trade — the fees. When you trade at high frequency, you pay the spread and the commission on every single round trip, and you pay them whether the trade was right or wrong. The strategy did not have to lose to lose money. It only had to be busy.

By the end of that month the retreat was explicit:

disable centers lower frequency lower frequency 2

Thirty-three days from TURN OFF LOCKING COMPLETELY to lower frequency.

The ugly

We never wrote down the reckoning. There is no post-mortem in the archive, no note explaining what the fees actually came to, no decision record saying high frequency was a mistake at our size and here is the number that proves it. There is only a quiet series of commits turning the dial back down, and then the subject never comes up again.

Which means that six years later, reconstructing this required reading commit messages one at a time and inferring the story from the order they were written in. The lesson survived in the code. The reasoning did not survive anywhere.

And there is a second admission owed here. Between the two ends of this arc sits a commit that says only:

the world goes round and round and the stars and the moons all balance each other

That is not an engineering note. That is somebody at the end of a long stretch, and it is in the permanent record of a production trading system. We are including it because a version of this story with only the tidy commits in it would be a different and less true story.

What we actually learned

Frequency is a cost multiplier before it is an edge multiplier. Every increase in trade count multiplies your fee drag immediately and your edge only if the edge is real. One of those is guaranteed and the other is a hypothesis. We had it backwards for about a month.

Turning off a safety to gain speed is a trade, so price it. last write wins is a perfectly respectable engineering choice in some systems. In a system that moves money it needs a written justification and a way to detect what it costs — neither of which we had.

The system that is busiest is not the system that is working hardest for you. This is the single most durable thing we took from 2020, and it is why the platform we run today deliberately trades a small number of instruments on a slow cadence, with position sizing and fee drag measured against broker statements rather than assumed.

Write down why you turned back. We are still paying, in reconstruction time, for a decision that was never documented — and this post exists because someone finally went and read the commits.

RoboTrader trades CME micro futures. Trading involves substantial risk of loss and is not suitable for every investor. Past performance is not indicative of future results.