Bitcoin forecasting is a crowded field. Dozens of models compete for attention — scarcity-based formulas, on-chain activity trackers, power-law charts, machine learning systems fed with macroeconomic data. And yet, for all that complexity, most of them can’t beat a simple guess.
That’s pretty much the finding from a preprint by Carlos Baquero from the University of Porto. After reviewing 23 significant studies on Bitcoin price prediction, Baquero found that no model consistently outperformed naive benchmarks over one-to-six-month periods. Naive benchmarks are basic — they just use today’s price or simple market trends, no fancy math. The paper is still awaiting peer review, but the pattern it lays out is hard to ignore. Sophisticated models, built by researchers with real resources, keep losing to the simplest possible baseline. That’s not a minor issue. That’s a structural problem.
Short-term forecasting is a different story, kind of. Order flow predictions and daily return models have shown some genuine value. But they’re often lumped in with longer-term price forecasts when people compare results — and that’s a mistake. Short-term and long-term prediction are basically different problems. Mixing them distorts the picture.
Why Markets Keep Breaking the Models
The core problem has a name: non-stationarity. Relationships between variables don’t stay stable over time. A model that worked well in one market era can become useless in the next. The 2017 Bitcoin cycle was retail-driven. The 2021 cycle was dominated by derivatives. Then spot ETFs arrived in 2024. Each shift changed the underlying dynamics enough to break models calibrated on earlier data.
Research by Francesco Puoti, Fabrizio Pittorino, and Manuel Roveri backs that up. Their study looked at naive models versus more sophisticated ones — ARIMA and LSTM networks — across multiple cryptocurrencies and forecast horizons. Naive models won. Consistently. The likely reason is that complex models tend to “memorize” noise in historical data rather than find real, stable patterns. They look impressive on past data. Then new data arrives and the whole thing falls apart.
And it’s not just a matter of model choice. Information leakage and sloppy data handling can inflate accuracy metrics to the point where they mean nothing. A model can show high historical accuracy and still produce forecasts that are useless in practice — forecasts you can’t actually trade on.
Backtest Overfitting and the Illusion of Success
Backtest overfitting is probably the single biggest trap in this field. Researchers keep refining models against historical data until the results look great. And they do look great — for the past. David Bailey’s research laid out exactly how this happens and why it’s so dangerous. The model fits the specific noise of a specific period. It doesn’t generalize. So when real future data comes in, the model fails.
The fix exists. Walk-forward evaluation and multiple holdout windows — methods that involve retraining models repeatedly on fresh data and testing them across varied market conditions — can catch overfitting before it becomes a problem. But Baquero’s review found that few of the 23 models he examined actually used these methods. Most didn’t. That’s a pretty damning finding for a field that presents itself as rigorous.
High accuracy numbers in backtests are seductive. They attract attention, get papers published, and sometimes move markets when they’re publicized. But if those numbers don’t translate into actionable trading strategies, they’re basically noise dressed up as signal.
Stock-to-Flow and Metcalfe’s Law Under Pressure
Two of the most popular Bitcoin valuation frameworks — stock-to-flow and Metcalfe’s Law — have their own specific problems. Stock-to-flow ties Bitcoin’s value to its scarcity, arguing that each halving reduces supply and pushes price up. Metcalfe-style models say Bitcoin’s value grows with its user base. Both are intuitive. Both have serious limits.
Alexander Shelton’s 2024 study tested Bitcoin return predictions using these frameworks. He found that stock-to-flow and Metcalfe variables could explain returns within the sample — the historical data they were built on. Outside the sample, that predictive power basically disappeared. The models fit the past. They didn’t predict the future.
Stock-to-flow has faced particularly visible pressure. Bitcoin’s actual price has diverged from the model’s projected path for extended periods. Supporters have responded by reframing the model — calling its predictions long-term value indicators or cycle averages rather than near-term price targets. That’s a reasonable adjustment, maybe. But it also makes the original claims much harder to evaluate. The goalposts moved.
Metcalfe’s Law runs into a different kind of trouble. Network activity and price tend to rise together, which looks like confirmation of the model. But they could both be rising because of adoption, or because higher prices attract more users, or because both are chasing some shared underlying trend. Establishing actual causation is really hard. Correlation isn’t enough.
Bitcoin has also gone through a limited number of independent market cycles. That’s a real constraint. Fewer cycles means less data to draw robust conclusions from, and models built on three or four cycles are probably fitting cycle-specific quirks rather than durable economic relationships.
Shelton’s study found that the tendency for these models to fit historical data well but stumble with new information keeps showing up across frameworks. No model has cracked it yet.
Hub: Bitcoin price, news, and analysis
Frequently Asked Questions
What did Carlos Baquero’s preprint find about Bitcoin forecasting models?
After reviewing 23 significant studies, Baquero from the University of Porto found that no Bitcoin price model consistently outperformed naive benchmarks over one-to-six-month forecast periods.
What is backtest overfitting and why does it matter for Bitcoin models?
Backtest overfitting happens when researchers refine models too closely against historical data, producing impressive past results that fail to predict future prices — a risk David Bailey’s research specifically examined.
Why It Matters
The findings from Baquero’s research highlight a fundamental challenge in Bitcoin price prediction that has implications for traders and investors alike. Despite the proliferation of sophisticated forecasting models, their inability to outperform simple benchmarks raises questions about the reliability of these methods in a market characterized by high volatility and uncertainty. This underscores the need for a cautious approach to Bitcoin investment strategies, as reliance on complex models may not yield the anticipated predictive advantage.