Experimental probability estimates an unknown likelihood from observed relative frequency. Its usefulness depends on both the amount of evidence and whether future trials resemble the trials you observed.
Separate the probability from its estimate
If an event occurs times in trials, the experimental estimate is The hat marks an estimate; it does not claim that the unknown probability is exactly this number. A result of 46 successes in 80 trials gives
A theoretical probability is derived from assumptions, such as equally likely outcomes. An experimental probability measures what happened. A discrepancy is a reason to examine both random variation and the assumptions, not automatically a reason to reject either the data or the model.
Example: predict a count and interpret it
A machine produces 7 defective parts in a random inspection of 250 parts. If production conditions remain comparable, the estimated defective fraction is Among the next 1,500 parts, the predicted defective count is
Forty-two is a forecast, not a guarantee or a maximum. It also assumes the inspection is relevant to the new production run. A larger sample from a different machine might be less informative than a smaller representative sample from this one.
If a calculation predicts 42.6 defective parts, do not conclude that a fraction of a physical part will be defective. An average or predicted count may be nonintegral even though each actual count is an integer. Round only when the requested report calls for a whole-number estimate.
Example: pool evidence correctly
A student makes 18 of 30 shots on Monday and 56 of 70 on Tuesday. The combined success rate is not the unweighted average of 60% and 80%, which is 70%.
The pooled calculation is a weighted average:
Tuesday receives more weight because it contains more shots. The estimate predicts 111 makes in 150 similar attempts. But if Monday's shots were long-range and Tuesday's were close-range, pooling them may not predict a new set of exclusively long-range shots well. Comparable conditions matter alongside arithmetic.
What a larger experiment does and does not show
Under stable, independent trials, a larger number of observations generally makes relative frequency more stable. It does not require the estimate to improve after every new trial, and it does not make the absolute difference between observed and predicted counts vanish.
Also, observing no successes is not proof that an event is impossible. Zero defects among 20 inspected parts is weaker evidence than zero among 20,000. The observed proportion is zero in both cases, but the amount of evidence differs greatly.
Avoid the gambler's fallacy: past independent outcomes do not make the opposite outcome due. An experiment supplies evidence about a probability; it does not create a balancing force on the next trial.
Check whether trials are comparable
Replacement can change an experiment. Drawing a colored counter and returning it before the next draw keeps the bag's composition fixed. Drawing without replacement changes the available colors. Before projecting a relative frequency, decide whether your future process follows the same rules as the observed process.
Try it yourself
A factory finds 3 flawed items in 50 inspected items and 9 in a second inspection of 150. What is the combined estimated flawed fraction, and the predicted count among 800 comparable items?
Answer: Pool the counts: The predicted count is Both inspections happen to have the same rate, but pooling remains the general method.
What to remember
Calculate a relative frequency, assess whether it is relevant to the prediction, and distinguish expected or predicted counts from certain outcomes. More representative evidence supports a more dependable estimate.