We're feeding astrology into a trading engine

We’re feeding astrological data into a trading engine. Before you close this tab, give us three minutes to explain why that’s not crazy.

The coastline problem

How long is the coast of Britain? It depends on your ruler. Measure with a 100km stick, you get one number. A 10km stick, a bigger number. A 1km stick, bigger still. Zoom in forever and the answer approaches infinity. This is Mandelbrot’s coastline paradox, and it’s not about geography — it’s about the fact that some problems get harder the more precisely you try to measure them.

Financial markets have this property. The signal space is fractal. More data, more features, more granularity — and you don’t converge on an answer, you drown in noise. Every quant who’s ever overfit a model has run face-first into this wall.

The solution to the coastline paradox isn’t a better ruler. It’s choosing a resolution and committing to it.

A clock with many hands

Set aside horoscopes and personality types for a moment. Look at what astrology actually is mechanically.

It’s a coordinate system. Ten or so celestial bodies, each orbiting at a different period — Mercury at 88 days, Jupiter at 12 years, Saturn at 29. The zodiac is just twelve 30-degree bins dividing each orbit. Aspects — conjunction, opposition, square, trine — are angular relationships between any two bodies. Houses add Earth’s daily rotation.

Strip away the mythology and you have a deterministic, high-dimensional system of overlapping oscillators at known frequencies, with a rich vocabulary of phase relationships between them.

That’s not mysticism. That’s a feature set.

Something like reservoir computing

In machine learning, there’s a technique called reservoir computing. You take a complex dynamical system — it could be a bucket of water, a neural network with random weights, anything with rich internal dynamics. You don’t train it. You just observe its state and train a simple readout layer on top.

The solar system isn’t a reservoir in the strict technical sense — a true reservoir receives inputs from the system you’re modeling, and the planets don’t care about your portfolio. But the principle is analogous: a complex dynamical system whose state you can read without training, providing a rich set of features for a lightweight model to work with.

Multiple interacting oscillators at different timescales. High-dimensional state that’s fully observable. Quasi-periodic orbits that rarely repeat on human trading horizons.

You don’t need to believe that Jupiter does anything to markets. You just need to ask: does the state vector of this dynamical system, read through the right projection, correlate with anything interesting in human behavior? And since markets are just human behavior with a price tag — it’s worth checking.

It might not. That’s fine. The question is testable, which makes it worth asking.

Why the grid matters

Here’s where most quantitative approaches to this idea would fail. You could compute the precise position of every known celestial body, calculate every pairwise angular relationship to ten decimal places, factor in velocity and declination and distance — and you’d have a feature space so large that any pattern you find is guaranteed to be noise.

More resolution doesn’t help. This is the coastline again.

Astrology solves this by being lossy. It says: these ten bodies, not the thousands of others. Twelve bins, not continuous angles. Six aspect types, not infinite angular relationships. This is dimensionality reduction — and the compression choices are the interesting part.

We want to be honest about where these choices came from. The ancient astrologers didn’t pick twelve bins because twelve was empirically optimal — they picked twelve because base-12 math was standard in Mesopotamia. They tracked those specific planets because those were the ones visible without a telescope. The grid’s resolution was set by the technological and cultural constraints of antiquity, not by some millennia-long feature selection process.

But that doesn’t close the question. It opens a different, more interesting one: did those constraints accidentally produce a useful decomposition? The bins are coarse enough to avoid overfitting. The feature set is small enough to be tractable. The phase relationships capture interaction terms that nobody in quantitative finance is constructing from scratch. Whether this is signal or coincidence is an empirical question — and we’d rather test it than assume the answer.

Three signal channels

If there’s anything in the grid, it could be operating through any of three mechanisms.

There are genuine physical effects. Lunar cycles influence tides, light levels, sleep patterns. Seasonal shifts affect mood, agriculture, energy demand, commodity supply chains. The astrological calendar captures some of these — particularly lunar and solar cycles — in ways the Gregorian calendar, designed for administrative convenience rather than celestial mechanics, handles clumsily.

There are culturally embedded rhythms. “Sell in May and go away” works not because May is cosmically special, but because enough people believe the pattern that it becomes self-reinforcing. School enrollment cutoff dates create measurable developmental differences between kids born days apart. Arbitrary temporal boundaries generate real behavioral effects. Astrology is a second temporal grid that’s been culturally embedded for thousands of years.

And there’s the reflexive layer. Financial astrologers exist. They trade on this. If enough of them act on the same signals, those signals move prices — regardless of whether the underlying mechanism is real.

For prediction, not needing to distinguish between mechanisms is pragmatically useful. But we should be transparent that mechanism-blindness has costs: it increases the risk of spurious correlation and makes the model fragile to regime changes. If the signal is purely reflexive and financial astrologers stop trading, the edge disappears. That’s a real limitation, not a technicality.

What we’re actually building

This isn’t a horoscope for your portfolio. It’s a hypothesis: that the astrological framework, treated as a feature engineering toolkit rather than a belief system, might contain signal that quantitative finance has ignored because the packaging is embarrassing.

Before we feed this grid into a live trading engine, we need to know whether it captures anything real about human behavior at all. Markets are driven by people. If the grid has no relationship to how influential humans cluster in time, there’s no reason to expect it to track the aggregate behavior we call prices. So the first test is deliberately upstream of markets.

The approach: take a well-known, defensible dataset — TIME Magazine’s 100 Most Influential People of the 20th Century — compute natal charts for each, build a null distribution from thousands of randomly sampled dates, and test whether any features in the grid are statistically overrepresented.

Two things to flag about the methodology.

The null distribution has to be built carefully. Births aren’t uniformly distributed across the year — there are seasonal patterns driven by conception rates, hospital scheduling, and cultural factors that have nothing to do with astrology. If we sample random dates uniformly, we’ll find “astrological” signals that are just seasonal clustering. The null needs to preserve marginal distributions for month, geography, and era, then randomize within those constraints.

And the data has limits. Birth dates for 20th-century figures are reliable. Birth times often aren’t — and the house system depends on exact time of day. That means houses may be unusable for a significant portion of the dataset. We’d rather drop a feature than fake the inputs.

If something survives multiple comparison correction with meaningful effect size, that’s interesting — and it earns the grid a seat at the table for the market application. If nothing does, we’ve falsified it cleanly against a dataset we didn’t curate. That’s also interesting, and we’ll publish it either way.

Every serious quant will tell you that the edge is in alternative data that others aren’t looking at. Nobody is looking at this — the reputational cost distorts research incentives even when the methodology is sound.

That might be the most interesting part.