What a prediction model is
A price prediction model is a set of rules, usually written as mathematics or code, that turns some inputs into a statement about a future price. The inputs might be past prices, trading volume, data recorded on the blockchain, survey results or news. The output might be a single number, a range, or a probability that the price will rise rather than fall.
Every model carries assumptions about how markets work. Those assumptions, not the sophistication of the maths, usually decide whether a model is useful. This article surveys the main families and what each one takes for granted. It does not make any forecast, and nothing here is financial, tax or legal advice.
Statistical time series models
The oldest family treats the price history as a series of numbers and looks for structure in it. Examples include autoregressive models, moving-average models and exponential smoothing. A close relative models the size of price swings rather than the direction. Such volatility models assume that calm and turbulent periods tend to cluster together.
The core assumption is that the past contains patterns that persist. For direction this is a strong claim, because researchers have often found very little reliable pattern in short-term returns. Volatility is easier to model than direction, which is why these tools tend to be used more for risk measurement than for calling tops and bottoms. The idea behind them is covered in more detail in the review of forecasting methods.
Chart-based and rule-based models
Many traders use technical analysis, which reads patterns in price and volume charts and applies rules such as crossovers of averages. When a rule is written precisely enough to be run on past data, it becomes a model of sorts: a mechanical signal with a definition.
The assumption is that crowd behaviour repeats in recognisable shapes. Whether it does is debated. A rule can look good on history by chance, which is why careful testing, sometimes introduced as backtesting basics, matters more than the rule itself.
Fundamental and valuation models
A different family tries to estimate what an asset is “worth” from underlying drivers. For a company this might be expected cash flows. For cryptocurrency there is no agreed equivalent, so people borrow ideas: usage of the network, the number of active participants, issuance schedules, or comparisons with other assets.
Some of these are explicit models. The stock-to-flow model, for example, relates scarcity to price and has attracted a lot of criticism. Others are looser frameworks built on supply and demand, asking what changes the number of buyers or sellers. The shared assumption is that price drifts toward some measurable anchor. Critics point out that the anchors are chosen after the fact and that a handful of past cycles is too little data to test them.
On-chain and sentiment models
Public blockchains record transactions, so analysts build indicators from them: the age of coins that move, flows to exchanges, or the cost basis of holders. The field is introduced under on-chain analysis. Sentiment models do something similar with a different input, using surveys, social media activity, derivatives positioning or composite measures of market sentiment.
These models assume that the behaviour of participants leaves a measurable trace and that the trace leads price rather than follows it. The second half of that assumption is the weak one. Many indicators move because the price already moved, so they describe the present better than they predict the future.
Machine learning models
Machine learning does not start from a theory of markets. It fits a flexible function to historical examples so that inputs map to outputs, and it can combine many of the data types above. Families range from simple regression to decision trees and neural networks.
The assumption is that the relationship between the inputs and the future price is stable enough to be learned from the past. Markets change as participants adapt, regulations shift and new products appear, so the stable relationship may not exist. Two articles in this cluster go further: one on how a model is trained, and one on where learning methods tend to fall short.
Scenario and simulation models
Some models do not try to name a price. Instead they simulate many possible paths from assumed rules and look at the spread of outcomes. This is common in risk work, where the question is how bad a loss could plausibly be, not what the price will be on a given date. The result is only as good as the assumed rules, and extreme events are often more frequent in real crypto markets than a simple simulation suggests.
Questions to ask of any model
Whatever the family, a few questions help separate a careful model from marketing.
- What does it assume about how prices behave, and is that assumption stated?
- Was it tested on data it had not already seen when it was built?
- How does it compare with a very simple baseline, such as assuming tomorrow looks like today?
- Does it give a range and an error rate, or only a confident-sounding number?
- Would it have been usable in real time, or does it rely on information that only became available later?
The article on evaluating predictions develops these checks, and the discussion of whether prices follow a random walk explains why the simple baseline is so hard to beat.
Why uncertainty does not go away
Cryptocurrency prices are volatile, and they respond to news, regulation, the wider macro environment and the behaviour of other traders, all of which are hard to anticipate. Even a model that fits history well can fail the moment conditions change. Past performance does not predict future results, and people can lose everything they put in.
The sensible way to use models is as tools for thinking about risk and for understanding what a claim assumes, not as machines that reveal the future. A model that is honest about its error rate is usually more valuable than one that sounds certain.