Thresholds and tipping points
Many systems change all at once when a line is crossed. A neuron fires, a classifier says yes, a language model drops its unlikely words, a price breaks out, a stop-loss sells. Each stop is a threshold, and each has to answer the same question: where to draw the line.
-
Pattern Essays
The Threshold
The pattern itself: a system stays quiet for a long time, then changes all at once when a level is crossed.
-
ML Math
Activation Functions
ReLU is a threshold. Below zero nothing passes, above it everything does, and stacking these switches is what lets a network bend.
-
The Toolkit
ROC & AUC Curves
A classifier outputs a score and a threshold turns it into a decision. The ROC curve shows every possible line at once, and what each one costs.
-
LLM Engineering
Sampling
Language models draw lines too: top-k and top-p sampling drop every candidate word below a cutoff before choosing the next one.
-
Chart Patterns
Support & Resistance
In prices, thresholds are levels traders remember: places where buying or selling has stepped in before.
-
Technical Indicators
Donchian Channels
A breakout system makes the line explicit: the highest high and lowest low of the last N bars, and a trade when price leaves that range.
-
Risk & Portfolio
Stop-Loss Strategies
Risk control uses the same switch on purpose: past a set loss the position is closed, without debate.
A threshold turns a continuous signal into a decision. Where you set it decides which mistakes you make, so every field ends up drawing its own ROC curve, whatever it calls it.
Add up enough small, independent nudges and the same curve appears — in sample averages, in price bands, in risk models. Follow it across four fields, then see where it quietly stops being true.
Extreme results tend to be followed by less extreme ones. It looks like a force pulling things back; it is arithmetic. Watch the same idea turn into trading signals, portfolio rules and model penalties.
Every field has the same problem: real structure mixed with randomness. These stops are the tools for pulling them apart — and for checking you have not invented a pattern that was never there.
When a system's output feeds back into its input, small nudges can snowball. Crowds, prices, volatility and even deployed models run on loops like this — and they tend to stay quiet until they don't.
The data you can see is often the data that survived. Patterns found in it can be an artefact of what went missing. These stops show how the trap works, and the habits that keep you out of it.
Recent direction tends to carry on, for a while. The same idea speeds up neural-network training, drives a family of trading indicators, shapes how traders add to winners and fuels the crowd. Follow it, and see where it stops paying.