Backtestify
Sign in
All articles
Forecasting Price ProbabilityNovember 9, 202614 min read

Forecasting Price Probability: A Practical Guide for Traders

Learn what a price probability forecast is, how it differs from a point prediction, and how to check it is calibrated before you risk real capital.

Forecasting Price Probability: A Practical Guide for Traders

Forecasting price probability means estimating how likely different future prices are, instead of naming one price and hoping. A useful forecast says something like "a 58% chance the price closes higher in five days, with a likely range between X and Y." You judge it by whether those percentages hold up on data the model has not seen, and by whether acting on them makes money after costs. This guide covers how these forecasts work, how to test them, and how to use them with a trading plan.

TL;DR

  • A probability forecast gives a range of outcomes and the likelihood of each. A point forecast gives one number and says nothing about uncertainty.
  • Common methods include Bayesian models, historical simulation, bootstrapped intervals, quantile regression and conformal prediction. Each has trade-offs.
  • Check calibration first: when a forecast says 60%, the event should happen about 60% of the time. Then check that acting on it pays.
  • Markets are noisy, so treat a probability as an input to risk sizing, never as a promise.
  • Backtestify's forecast tool shows probability and confidence statistics for a price move, and you can test your rules on real history first.

Table of Contents

What a probability forecast actually is

A point forecast predicts a single value, such as "the stock closes at $104 on Friday." A probabilistic forecast describes a range of possible values and how likely each one is, while a point forecast puts all of its weight on one number and carries no description of uncertainty.

The practical difference is large. Point forecasting tells you about the expected value, for example tomorrow's return. It does not tell you how uncertain that estimate is. Quantiles or intervals add information about unexpected fluctuations, which supports better decisions. Two forecasts can share the same central guess of +1%. One may say the outcome sits between +0.8% and +1.2%. The other may say it sits between -6% and +8%. You would size those trades very differently.

For a trader, a probability forecast usually takes one of three shapes:

  1. A direction probability. The chance the price is higher (or lower) after a set horizon.
  2. A range or interval. For example, an 80% interval for where the price will be in ten days.
  3. A full distribution. The likelihood of every outcome, from which you can read any percentile.

The first shape is the easiest to test. The third carries the most information. Most retail decisions need only the first two.

How price probabilities are estimated

There is no single method. The research literature groups the common ones into a handful of families. A 2026 review of probabilistic price forecasting lists Bayesian models, historical simulation, bootstrapped prediction intervals and quantile regression among the standard approaches. It also notes that they can struggle when the data's behavior changes over time and often lack formal coverage guarantees.

Historical simulation and bootstrapping

The simplest approach assumes the future will resemble a resampled version of the past. You take past returns, reshuffle or resample them thousands of times and read the spread of results. It is intuitive and needs few assumptions. It is also only as good as the history you feed it, which is why the period you pick matters.

Distribution assumptions and indicators

Some traders assume returns follow a lognormal distribution and convert a target move into a Z-score. Others use technical indicators such as moving averages, RSI or Bollinger Bands as a rough proxy for odds. Option pricing models go further and bake volatility and time to maturity into the estimate. These methods are quick to compute. Their weak spot is that real returns have fatter tails than a tidy bell curve suggests, so rare large moves happen more often than the formula implies.

Machine learning models

Machine learning models learn from historical data to estimate future prices. They can ingest market trends, economic indicators and news sentiment, and they can keep adapting as new data arrives. The same flexibility lets them fit noise, so they need disciplined out-of-sample testing.

Conformal prediction

Conformal prediction wraps a standard model in a calibrated interval. In one practitioner description, it works in three steps: train a baseline point model, calibrate on a separate dataset using the residuals, then test by overlaying the residual distribution on the point forecast. It is attractive because it can give valid intervals under light assumptions. Its standard guarantees rely on exchangeability, and that assumption is generally violated in time series, so adaptive versions matter for price data.

Three steps for calibrating a forecast interval

Futures prices as a forecast

For commodities, futures prices are a built-in market forecast. A 2026 Federal Reserve note found that futures can outperform several alternative approaches when the slope of the futures curve is meaningfully different from zero. The same note says futures prices do not in general produce unbiased forecasts, which it links to hedging, inventory dynamics and illiquidity. Read them as one informative input, not a final answer.

Three-way forecasts: up, down, or stay out

Most direction forecasts force a choice between up and down. A three-way (ternary) forecast adds a third class for relatively small moves in either direction. Research on trading via selective classification describes this as giving the model an extra way to avoid calling a direction.

The same paper explains why this suits trading. A selective classifier can abstain and stay flat when it is not confident enough. In the ternary version, the strategy can stay flat either when the predicted label is the small-move class or when the classifier abstains.

This maps neatly onto how disciplined traders already work. Not every day deserves a position. A forecast that says "54% up, 46% down" is close to a coin flip, and a rule that skips those days can improve the quality of the trades you do take. You can test that exact filter in a backtest by adding a minimum probability threshold to your entry rules.

How to read a probability forecast

Three ideas cover most of what you need.

Calibration

Calibration asks whether the numbers mean what they say. Formally, conditional on any forecast value p, the event should occur in 100·p% of the instances. If a model gave 60% on 200 separate occasions, the price should have risen in roughly 120 of them. A reliability diagram plots the stated probability against the observed frequency, and deviations from the diagonal suggest miscalibration.

For interval forecasts, the same logic applies: if a 20% quantile is predicted, about 20% of observations should fall below it.

Sharpness

Sharpness is how narrow the predicted distribution is. A narrower range implies higher confidence. Sharpness alone is worthless, though. A forecast that is narrow and wrong is worse than a wide one that is honest. You want the sharpest forecast that is still calibrated.

Economic value

Statistical quality does not guarantee profit. The 2026 review cautions that trading profit or downside-risk reduction is a complementary evaluation and does not automatically follow from better calibration or coverage. A well-calibrated 51% edge can vanish after spreads and commissions. So you test both layers.

CheckQuestion it answersWhat good looks like
CalibrationDo stated probabilities match observed frequency?Points close to the diagonal on a reliability plot
CoverageDoes a 80% interval contain the outcome about 80% of the time?Empirical coverage near the stated level
Interval widthHow tight is the range?Narrow without losing coverage
Scoring ruleWhich forecaster is better overall?Lower Winkler score or CRPS, or pinball loss where relevant
Economic resultDoes acting on it pay?Positive expectancy after costs, with acceptable max drawdown

The first four rows use metrics the review lists for probabilistic evaluation: empirical coverage, average interval width, Winkler score, CRPS and pinball loss. The last row is where a backtest comes in.

A worked example

This example is illustrative, with made-up numbers to show the arithmetic.

Suppose a model says there is a 60% chance an index closes higher in five trading days. Your rule: buy when probability is at least 60%, risk 1R (one unit of risk) on a stop, and target +1R.

  • Over a long sample, the model issued 60% or higher on 200 occasions.
  • The price actually rose after 120 of them, so the forecast is calibrated at that level.
  • If each win pays +1R and each loss costs -1R, expectancy is 0.6 × 1 − 0.4 × 1 = +0.2R per trade before costs.

Now add a cost of 0.1R per round trip. Expectancy drops to +0.1R. Add slippage on a fast market and it can fall further. A forecast can be perfectly calibrated and still be marginal once you price in execution. That is why the economic check is separate.

Now suppose the same model produced only 108 up moves in those 200 cases (54%). It was overconfident. Your realized expectancy is 0.54 − 0.46 = +0.08R before costs, which can turn negative after them. The forecast number on screen looked the same. Only testing exposes the difference.

A worked example

Why a forecast can miss the next move

Even a well-built model will be wrong often, and that is expected from a probability. A 60% forecast loses 40% of the time by definition. The deeper question is whether there is any signal at all.

Financial data has a low signal-to-noise ratio. One NeurIPS invited talk argues that a share price depends on anticipated future conditions rather than only its own past, and that efficient markets limit predictability beyond a naive forecast. Recent research on news-based prediction points the same way: in one 2026 study of zero-shot models, no model exceeded the majority-class baseline of 48.4% for short-horizon stock movement from news sentiment alone.

The field itself is still maturing. A 2025 survey reports that probabilistic AI for financial time series faces a lack of standardized benchmarks and evaluation metrics, limited financial interpretation and a fragmented literature. The takeaway for you is practical: demand evidence on out-of-sample data, keep position sizes proportional to the uncertainty, and treat each probability as one input to a risk-managed plan.

Test probabilities before you risk capital

A forecast earns trust through a repeatable process. Here is one you can run.

  1. Write the rule in plain terms. Define the entry, exit, stop and the probability threshold. For a refresher, see our step-by-step guide to backtesting a trading strategy.
  2. Backtest on real historical data. Record win rate, profit factor and max drawdown, not just total return.
  3. Split your data. Fit on one period and evaluate on another. Walk-forward analysis repeats this across rolling windows and helps you avoid overfitting.
  4. Stress the result. A single equity curve is one path. Monte Carlo backtesting reshuffles outcomes over thousands of paths to show tail risk and a realistic worst case.
  5. Check calibration of the probabilities. Bucket forecasts (50 to 55%, 55 to 60%, and so on) and compare each bucket to what actually happened.
  6. Compare against a baseline. Judging a forecast without a yardstick is easy to get wrong. Choosing a benchmark that fits your strategy gives you something fair to measure against.
  7. Forward test. Run the rules on new data as it arrives, using the forward-testing checks before committing larger size.

If you trade shorter horizons, the intraday backtesting guide covers the extra care that fast data needs. If you hold for days, the swing trading backtest checklist is the closer fit.

Putting probabilities to work in Backtestify

Backtestify is built for the verification step. You enter a strategy in plain English, run it on real historical market data and see win rate, profit factor and max drawdown trade by trade. You can improve the rules, compare the original and improved versions side by side on recent data, and export the result as TradingView Pine script.

The Forecast tool adds probability and confidence statistics for the next price movement, so you can see how likely a move is and judge it alongside your tested rules. We also publish results for winning and losing strategies, so you can see how well-known setups held up. One example is the Turtle Trading System 1 report.

The Free plan lets you start at no cost. Pro, which unlocks unlimited backtesting, improvement and forecasting, is $29 per month or $190 per year. See the plans on the Backtestify home page.

Backtestify

If you are not ready to subscribe, start with the Free plan and run your first test.

Common mistakes when using price probabilities

  • Treating 70% as certainty. It loses three times in ten. Size so a losing streak is survivable.
  • Judging on a handful of trades. Calibration needs many forecasts per bucket before the observed frequency means anything.
  • Tuning until it fits. Every extra parameter you adjust against the same data raises the odds you are fitting noise.
  • Ignoring costs. Spreads, commissions and slippage come out of a thin edge first.
  • Using one market regime. Test across rising, falling and sideways periods.
  • Confusing a tight range with accuracy. Narrow intervals only help when coverage holds.

A useful habit is keeping a simple log: date, stated probability, outcome. After a few hundred entries you can build your own reliability table without any special software.

FAQ

What is a probability forecast?

A probability forecast states the likelihood of outcomes instead of naming one. It might give a 62% chance of a higher close, or an 80% interval for where price will land. Unlike a point prediction, it expresses uncertainty, which helps with position sizing and risk limits.

What is three-way forecasting?

Three-way forecasting adds a "small move" class to the usual up and down. The model can then decline to call a direction when the move is likely to be small or the signal is weak. In trading this means staying flat on low-conviction days, as described in research on selective classification.

How do I know if a price probability is reliable?

Compare stated probabilities with what happened. If forecasts of 60% came true about 60% of the time on enough samples, they are calibrated. Then run a backtest to confirm that acting on them gives a positive result after costs, and confirm it on data the model did not train on.

Can anyone predict the next price movement?

No method gives certainty. Research points to low signal-to-noise in markets and limited predictability in efficient ones. The realistic goal is a small, testable edge that you manage with sound risk control, and a probability is a tool for sizing that edge.

How do I forecast a share price probability myself?

Start with historical returns and estimate how often the price rose over your horizon. You can refine with volatility, indicators or a model. Then validate it with a backtest and a walk-forward test before using real money.

Sources

Recommended