People will tell you their cat predicts the weather. Mine apparently does: before a storm she gets restless and goes under the bed, and I have caught myself checking the sky because of it.
Here is what's actually happening. She goes under the bed several times a week, for reasons including thunder, the hoover, a delivery van, and a mood. When a storm follows, I remember it. When no storm follows, I don't. The hit rate is unimpressive and the story is excellent, because I am the one keeping score and I am not trying very hard.
Every popular bitcoin price model works this way. Not because the people building them are dishonest — mostly they aren't — but because the structure of the problem manufactures convincing patterns out of very little data.
1. The Models You've Seen
- Stock-to-flow. Values bitcoin by the ratio of existing supply to new issuance, producing a stepped line that rises at each halving. It fitted the early history beautifully and was, for a while, the most cited chart in the space. Its published projections then diverged badly from what actually happened, and the model's defenders spent a cycle explaining why.
- The rainbow chart. A logarithmic regression through the price history with coloured bands above and below, labelled from "fire sale" to "maximum bubble". The bands are drawn to look right. There is no mechanism underneath — it's a trend line with decoration.
- Power law fits. More serious, and genuinely interesting: price plotted against time on log-log axes does produce a striking straight line. The open question is whether that reflects something structural about network growth or whether any young exponential-ish thing looks like this until it doesn't.
- Cycle-timing models. Count months from the halving, or from the previous top. Covered in my halving article: three or four observations is not a dataset.
2. Why They're So Convincing
Three mechanisms, and they apply to all of the above:
Overfitting is easy with sixteen years of data. With few enough observations, many different curves fit the past well, and they diverge wildly about the future. Fitting history is not evidence — it's the minimum entry requirement, and it costs nothing.
Survivorship does the marketing. Dozens of models get proposed each cycle. Two or three happen to match what follows, and those are the ones you hear about. The failures aren't refuted, they're forgotten, so the field looks far more successful than it is.
Log charts flatter everything. On a logarithmic axis, a model that's off by 300% looks like a small vertical gap. Charts that look like a tight fit are frequently describing errors large enough to have ruined anyone trading them.
Add the fact that a confident number is more shareable than an honest range, and the incentives point entirely one way.
3. The Deeper Problem
Even a well-built model faces something that doesn't apply to physics: the thing being modelled reads the model.
If a projection becomes widely believed, people position for it, and the positioning changes the outcome. A model that everyone trusts stops describing the system and starts perturbing it. Markets are reflexive in a way that weather is not — the sky doesn't behave differently because a cat went under the bed.
And underneath that: bitcoin's price is set by the marginal buyer, whose behaviour depends on liquidity conditions, regulation, technology, and human mood. No supply-side model can capture those, because supply is the one part of this system that was never the uncertain bit. We have always known exactly what the supply will do. That's not where the question lives.
4. What's Actually Usable
I'm not saying throw away every framework. Some things do carry information, as long as you hold them at the right strength:
- Long-horizon ranges over point forecasts. "Probably somewhere in a wide band over ten years, with large drawdowns along the way" is defensible. A price and a date is not.
- Structural indicators as context, not signals. Supply in long-term storage, exchange flows, fee revenue. They describe how the room is set up. They don't say when.
- Explicit falsification. The single most useful question about any model: what would we see if this were wrong? Anything that can absorb every outcome is a story.
- Your own plan as the primary instrument. A schedule, a size, and a written exit beats every model, because it works regardless of which one turns out to be right.
5. The Test to Apply
When you next meet a chart with a confident forward line, ask three things:
- How many independent observations is this fitted to? In bitcoin the honest answer is almost always "three or four cycles", which is not enough for the confidence being displayed.
- What's the mechanism? Why should this relationship hold? "It has so far" is a description, not a reason.
- Who benefits from you believing it? Usually not you.
Three questions, thirty seconds, and most of what circulates in this space resolves itself.
The Point
My cat does not predict the weather. She is a creature with moods and excellent hearing, and I am a pattern-matching animal who remembers the hits. The system that produces the impression of prophecy is entirely on my side of it.
Bitcoin price models are that, dressed as mathematics. The sixteen-year history is real and the curves are real; the confidence is manufactured by how few storms we've actually seen, and by who does the remembering.
Anything that fits every outcome predicted none of them. 🐾⚡
