Volatility can be forecast better than the standard method, but sizing by the forecast added little
Volatility is the one thing in markets that can be forecast with some skill. I wanted to know how much a neural network or gradient boosting adds to the standard method, and whether sizing trades by the forecast is worth it. As before, the pass criteria were written down before the run.
Setup
- XAUUSD H1 (HistData, 2010 to 2025) and USDJPY H1 (OANDA, 2010 to 2026.06). Test years 2014 to 2025 for gold and 2015 to 2025 for USDJPY, each trained only on the years before it (the last 5 days removed).
- Target: the average true range of the next 24 H1 bars, divided by the current ATR(14) (log). Forecasting "the current ATR continues" is the persistence baseline.
- Methods: persistence, HAR (the standard linear model on 1, 24, 120 and 480 bar ranges), LightGBM and a small neural network (MLP, three seeds averaged). Accuracy: QLIKE, lower is better.
- Pass criteria, fixed in advance: at least 5% better than persistence, at least 2% better than HAR, and better than HAR in at least two thirds of the test years.
Forecast accuracy (QLIKE, all test years)
| Symbol | Persistence | HAR | LightGBM | MLP |
|---|---|---|---|---|
| XAUUSD | 0.0315 | 0.0244 | 0.0218 | 0.0222 |
| USDJPY | 0.0392 | 0.0334 | 0.0291 | 0.0296 |
All four combinations passed all three criteria. LightGBM beat HAR by about 11% on gold and 13% on USDJPY, and it was better than HAR in every test year. So volatility, unlike direction, can be forecast better than the standard method does.
Does it help to size trades by the forecast?
To isolate the sizing effect, I used the hourly returns of holding one unit of the symbol and multiplied them by (reference ATR / forecast range), limited to 0.3 to 3 times.
| Symbol | Sizing | Month-to-month variation of realised volatility | Maximum drawdown | Sharpe |
|---|---|---|---|---|
| XAUUSD | constant | 0.308 | 0.280 | +0.75 |
| XAUUSD | persistence (ATR) | 0.224 | 0.410 | +0.75 |
| XAUUSD | LightGBM | 0.223 | 0.341 | +0.80 |
| USDJPY | constant | 0.373 | 0.231 | +0.26 |
| USDJPY | persistence (ATR) | 0.266 | 0.223 | +0.30 |
| USDJPY | LightGBM | 0.263 | 0.224 | +0.32 |
Sizing by volatility makes the month-to-month variation about 25 to 30% smaller than a constant size. But the plain ATR does almost the same: LightGBM added 0.223 against 0.224 on gold and 0.263 against 0.266 on USDJPY. Two of my criteria for the application failed: the variation had to be at least 5% smaller than with HAR, and on gold the maximum drawdown had to be smaller than with constant size (it rose from 0.28 to 0.34, because the size grew in quiet periods and a sudden move followed; it depends on the cap of 3 times).
What I take from it
- A better forecast is not the same as a better result. Most of the benefit of volatility sizing is already in the simple ATR rule.
- Scaling the size down when volatility is high is cheap and useful. Scaling it up when things are quiet raises the risk of a drawdown after a sudden move, so the cap matters.
- One data set per symbol, one definition of the target. Other targets, features or horizons may behave differently.
Past data only; nothing here is a forecast or advice.
