Calibrating Directional Signal Confidence with a Statistical Model

Summary

The document asks whether the strength of a technical indicator, such as RSI, can be converted into a probability that an asset will move in a given direction. The answer cautions that indicator readings alone do not supply a calibrated likelihood. A model must specify an underlying signal, such as genuine overpricing, and distinguish it from noisy observations.

Within that framework, repeated observations can help reveal the latent signal by averaging away noise, using large-sample behavior or a central limit theorem argument. The response gives a general modeling principle rather than a specific estimator, probability mapping, or trading rule. It does not identify a dataset or show empirical evidence, and any confidence estimate would depend on the model’s assumptions about the signal and noise distribution.

Key ideas

• An indicator’s magnitude does not by itself provide a probability for a future price direction.
• A confidence estimate requires a model of the underlying signal and the noise that obscures it.
• Averaging observations can help separate persistent signal from random variation under suitable assumptions.
• Large-sample results can support inference, but the response specifies no concrete calibration procedure.
• Any resulting probability depends on the validity of the assumed model.

Tags

# Determining confidence level of directional signals


# Determining confidence level of directional signals












With regards to technical analysis, are there ways of determining the confidence level of a directional signal? Taking a relative strength index (RSI) as an example, can the extent to which an asset is oversold or overbought be used to get the percentage likelihood of a movement in either direction?

## Answer by Hans (score 0, accepted)

https://quant.stackexchange.com/a/25838

You need a model which assumes that some intrinsic properties such as true overpricing is taking place but masked by the noise which has some probabilistic distribution around the true signal. The averaging of the observed data then takes advantage of the large number theorem or some version of the central limit theorem to flush out the signal. So you need to build a model first.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.