What is HFT Tracker?
Mapping low-latency infrastructure.
High frequency trading is commonly described in software terms. However, most of it is actually a civil engineering problem: towers, line of sight, and the seven hundred miles of air between two data centers.
CME Group matches futures in a data center in the suburb of Aurora, outside Chicago. Prices set there lead the equities and ETFs traded at the New York venues in New Jersey: Secaucus, Carteret, Mahwah. A firm that sees the Aurora price a few thousandths of a second before the rest of the market knows something about New York that New York has not priced in yet. All of this infrastructure investment is in the pursuit of that advantage.
Why air vs fiber?
Light moves through glass roughly a third slower than it moves through air. Wherever a clear line of sight exists, a chain of microwave dishes soundly beats the best fiber over that same route, at the cost of bandwidth and of a link that heavy rain will degrade. Millimeter wave carries more, but only over short hops, so it shows up clustered around the endpoints rather than spanning the corridor.
Every path on the map is that trade-off made physical. Each one ends at a real antenna, on a real tower or rooftop, drawn at the height the FCC has on file.
Where the data comes from
Everything on the map is drawn from two public records.
| Source | What it contributes |
|---|---|
| FCC ULS | Licensed point-to-point microwave links, matched on the contact e-mail domain filed with the licence, which is the practical way to catch an operator trading under a name that gives nothing away. The list of entities to watch was first compiled by Professor Lariviere at UIUC. |
| Micronet March 2025 |
Registered millimeter wave paths, unfiltered: at these distances the band is used almost exclusively by traders. The FCC publishes no bulk file for this band, so the layer comes from the registry itself and is refreshed by hand rather than weekly. |
What the pipeline corrects
The FCC file is a filing record, not a survey, and it is wrong in specific and repeatable ways. Each weekly rebuild fixes what it can prove and declines to guess at the rest.
- Elevations filed in feet. The ground elevation field means metres, and some filings put feet in it, which hangs the antenna hundreds of metres above the rooftops. I found it as a hop floating over a building in downtown Chicago. Catching one is easy: a single site cannot have two ground elevations that differ by exactly the feet-to-metres conversion, so one of them is a unit mistake. Deciding which one is the hard half. The obvious rule, that the bigger number is the feet, is wrong often enough to do real damage, because on high ground it is the smaller number that was botched. Every disputed site is checked against a real elevation instead.
- Blank elevations. A missing ground elevation is never filled in with zero, which would drop the antenna to sea level. The link is flagged and drawn relative to terrain instead.
- Dead licences. Drawn unfiltered, most of the paths in the file are cancelled, expired or terminated. The map shows active licences by default and draws expired ones dashed, since a network that was built and then given up is still worth seeing.
- Duplicate hops. A dual-polarised link is filed as two paths over identical geometry. They collapse into one line, with the polarisations kept as attributes.
- Name variants. “New Line Networks LLC” and “New Line Networks, LLC” are one operator, and are grouped as one.
- Curvature. A hop is a straight line through space, but interpolating it in latitude and longitude bows it above the true ray, and the longer the hop the worse it gets. Sampling happens in cartesian coordinates, so the line you see is the ray.
Where nothing can settle a disagreement, the record is drawn exactly as filed. Heights are the FCC’s own figures above mean sea level with the geoid offset left in, so altitudes are self-consistent rather than surveyed.
What the map could still reveal
- Latency technique. Hop spacing, antenna heights and choice of band are all decisions someone made, and reading them closely is reading the microseconds an operator was trying to shave.
- A latency floor. Path geometry plus an estimate of equipment delay gives a lower bound on how fast a message can physically cross the corridor.
- Reaction time. Compare that floor against the observed gap between a move in Aurora and an order arriving on a New York book, and you get a bound on how quickly a specific firm actually reacts.
Why firm names are hard to pin down
That last step needs a path tied to a firm, and it is where the public record stops cooperating. A proprietary network is a competitive asset, so firms go to real lengths to obscure it, often licensing under a name with no visible connection to trading. Attribution is probably not solvable from communications data alone.
It also cuts the other way: the published files carry stable operator labels rather than call signs and licensee names, since those are lookup keys straight back into the FCC record and would hand over the filter list itself. The paths are public either way. It is the curation that is worth protecting.
The exception is the firms that sell access rather than trade on it. McKay Brothers is a well-known third-party provider, and its paths are easy to find in the FCC data, because advertising the route is how it wins clients. Firms trading their own capital have precisely the opposite incentive.
Spot something wrong, or a link that should be here and isn't? Send it over. I read every suggestion by hand.