I’ve been looking at the growing number of Bitcoin price-prediction models, and one thing stands out to me: many of the most complicated systems still struggle to beat a very simple forecast. Whether the model uses power laws, on-chain data, macro indicators, or artificial intelligence, complexity does not automatically mean better predictions.
Bitcoin has inspired hundreds of forecasting methods. Some models use the halving schedule and scarcity to estimate future value, while others rely on wallet activity, transaction data, or network growth. More advanced systems use machine learning to process market and macroeconomic information.
But all of these models face the same basic competitor: a "naive forecast."
A price forecast can simply assume tomorrow's price will be close to today's price. A return forecast can assume the next return is zero, while a direction forecast can effectively follow a random walk. Surprisingly, sophisticated models often struggle to consistently outperform these basic approaches when tested on new market conditions.
A May 2026 preprint by Carlos Baquero of the University of Porto reached a particularly interesting conclusion. After reviewing Bitcoin forecasting research, the study found that no model had demonstrated durable superiority over an appropriate naive benchmark across multiple market regimes and one- to six-month forecasting horizons.
The review examined hundreds of papers but selected 23 for closer analysis based on factors such as methodology, influence, and genuine out-of-sample testing. The paper is still awaiting peer review, so its conclusions should be treated accordingly. Still, the central message is important: forecasting models need to prove that they can work beyond the historical data used to create them.
The Problem With Complex Models
I think the biggest problem is not necessarily that these models are badly designed. The bigger issue is that Bitcoin's market keeps changing.
A strategy that worked during the retail-driven 2017 cycle faced a very different market in 2021, when derivatives became more important. The arrival of spot Bitcoin ETFs in 2024 created another major channel for capital and price discovery.
This is known as "non-stationarity." In simple terms, the relationship between different variables can change over time.
Bitcoin's liquidity, investors, regulation, market access, and trading infrastructure have all evolved. A model can discover a relationship that looks extremely powerful during one period and then fail when the environment changes.
Research by Francesco Puoti, Fabrizio Pittorino, and Manuel Roveri found something similar. Their study compared 12 statistical, machine-learning, and deep-learning methods across five major cryptocurrencies at one-day, seven-day, and 30-day horizons.
The simpler models consistently performed better than methods including ARIMA, Prophet, random forests, XGBoost, LSTM networks, and N-BEATS.
That doesn't mean artificial intelligence or machine learning is useless for crypto. It means a complicated model only creates an advantage when there is a stable pattern for it to learn. When the underlying signal is weak or temporary, the model can simply "memorize the noise."
And Bitcoin has plenty of data that can make this problem worse. Millions of hourly or minute-level observations may look like an enormous dataset, but many of those observations come from the same market regime. Repeating thousands of observations from a single bull market or liquidity shock doesn't necessarily give a model thousands of independent lessons.
When Backtests Start Looking Like Crystal Balls
Another issue I find important is "backtest overfitting."
Researchers can test different variables, time periods, indicators, lookback windows, and model architectures. Eventually, one version is likely to produce an impressive historical result simply by chance.
That winning model may have discovered something real. But it may also have won what is essentially a "lottery" conducted on historical price data.
David Bailey and his co-authors studied this problem and showed that testing more variations increases the probability of finding an impressive backtest even when the underlying strategy has little genuine predictive power.
This is why a single historical test isn't enough.
Walk-forward testing provides a stronger approach because the model repeatedly learns from past data and then makes predictions on periods it hasn't seen. Even better, researchers can use multiple non-overlapping holdout periods so the model has to survive different conditions, including bull markets, crashes, sideways markets, and changing liquidity.
Information leakage creates another problem. If future information accidentally enters the training process, the model can appear far more accurate than it really is.
Even the performance metric can sometimes create a misleading impression. Bitcoin prices are persistent, so a model that predicts $100,500 when Bitcoin actually moves from $100,000 to $99,500 may have a relatively small price error while still producing the wrong trading signal.
For traders, direction, magnitude, timing, and transaction costs matter much more than simply being close to the eventual price.
The Models That Sound Better Than They Forecast
Some of Bitcoin's most famous valuation models remain popular because they provide simple explanations for a complicated market.
Stock-to-flow argues that scarcity can drive value, with each Bitcoin halving reducing the amount of new supply relative to existing supply.
Metcalfe-style models connect network value with the size or activity of the user base.
Power-law models attempt to describe Bitcoin's long-term price trajectory through a mathematical relationship between price and time.
I don't think these ideas are automatically meaningless. They can provide useful frameworks for understanding Bitcoin's history. The problem begins when historical relationships are treated as reliable future forecasts without enough out-of-sample evidence.
A 2024 peer-reviewed study by Alexander Shelton found that stock-to-flow and Metcalfe-related variables could help explain Bitcoin returns inside the sample, but their predictive power outside the sample was limited or disappeared.
The stock-to-flow relationship becomes especially interesting because Bitcoin's supply ratio increases according to a predetermined schedule, while Bitcoin's price also increased dramatically during much of its history. Two variables moving together through time can create the appearance of a strong economic relationship even when the underlying causal connection is weaker than it looks.
Metcalfe-style models face a similar challenge. Network activity can increase because Bitcoin adoption is growing, but higher prices can also attract more users and activity. Both variables can therefore influence each other.
Why Power Laws Are Different
Power-law models are more complicated to dismiss because their long-term curves have tracked significant parts of Bitcoin's historical journey.
They can be useful as a visual framework for showing whether Bitcoin is trading above or below a long-term trend.
But a high "R-squared" value does not automatically prove that the underlying relationship will continue into the future. A model can fit historical data extremely well and still fail when new observations arrive.
Researchers need to examine whether the results remain stable when the starting date changes, whether alternative mathematical functions perform similarly, and whether the model actually works on future data.
That's where the difference between "describing the past" and "predicting the future" becomes critical.
What a Better Bitcoin Forecast Should Look Like
For me, the most useful standard is actually quite simple.
A forecasting model should show its performance next to a naive benchmark. Researchers should test it across different market regimes, include realistic trading costs, and make the data and code available for independent verification.
They should also disclose how many different versions of the model were tested before presenting the winning result. Without that information, it's difficult to know whether the reported performance represents genuine predictive power or simply the best result from hundreds of experiments.
Most importantly, valuation models should not automatically be presented as precise price forecasts.
Sometimes the most honest conclusion is that "today's price is the best forecast."
That answer isn't exciting. It doesn't provide a huge price target or a specific date for Bitcoin to reach it. But it does something many complicated forecasts fail to do: it forces the model to prove exactly how much additional information it contributes beyond what the market already knows.
And that, in my view, is the real test for any Bitcoin prediction model.
#bitcoin #BTCReaches$80000 #BitcoinETFsBiggestDailyInflowSinceJanuary