Regime detection in seconds, not minutes
Indicators like ADX take a while to react. By the time they confirm a trend, much of it is over. For scalping, regime detection needs to be faster.
Alternatives use the data available at tick scale. Tick rate shows activity. The ratio of net price change to total distance traveled over the last minute, sometimes called the efficiency ratio, shows whether movement was directional or choppy. A value near one means a clean trend. Near zero means chop.
Autocorrelation of short-term returns hints at momentum or mean reversion. Spread behavior adds another dimension: widening spreads often accompany directional pressure.
The risk is overfitting. Regime labels are imprecise, and a model that switches too eagerly produces whipsaws.
Testing out of sample is essential. The regime detector should help in data it wasn't tuned on, otherwise it's decoration.
Trending or ranging in the last minute: which measure decides it?
Regime detection in seconds, not minutes
Regime detection in seconds, not minutes
It’s Fairman 
-
FXS HongKong Scalper
- Posts: 543
- Joined: Sat Oct 24, 2026 10:15 pm
Re: Regime detection in seconds, not minutes
The efficiency ratio is a neat choice because it costs almost nothing to compute. One thing worth deciding up front is the lookback in ticks rather than seconds, otherwise a quiet Asian hour and a busy London minute produce very different readings for the same market behaviour.
FXS Official Team | Hong Kong Session Desk
-
Amanhasnoname
- Posts: 2060
- Joined: Fri Oct 30, 2026 9:45 am
Re: Regime detection in seconds, not minutes
For me the thing that finally worked was a plain count of direction changes rather than anything clever. I take the last two hundred ticks, count how many times the mid price flips from up to down or back, and compare that with the same window on a normal day for that hour. When the flips drop well below the usual level the market is moving in one direction, and when they spike I treat it as chop and sit on my hands.
I tried the efficiency ratio first and liked it, but on its own it flickered too much around the middle values. What calmed it down was requiring the reading to stay on one side for about thirty seconds before I act on it, and ignoring the measure completely in the five minutes either side of scheduled news.
The overfitting warning is fair. My first version had four inputs and looked brilliant on the month I built it on, then did nothing useful the following month. Two simple inputs with loose thresholds have held up far better for me.
I tried the efficiency ratio first and liked it, but on its own it flickered too much around the middle values. What calmed it down was requiring the reading to stay on one side for about thirty seconds before I act on it, and ignoring the measure completely in the five minutes either side of scheduled news.
The overfitting warning is fair. My first version had four inputs and looked brilliant on the month I built it on, then did nothing useful the following month. Two simple inputs with loose thresholds have held up far better for me.