Forecasting cryptocurrency prices using high-frequency data
A good cryptocurrency forecasting model that has worked for me is taking linear combinations of technical indicators. The technical indicators represent the momenta of market-microstructure effects - order imbalance, trend-following, etc. It’s lightweight and not computationally intensive so it can be run on cheap hardware.
A future price of a cryptocurrency is computed using the current price plus the summation of weighted technical indicators.
The predictive power of this model comes from:
- Calculating and continuously updating the coefficients of the indicators in real-time and accurately scaling the coefficients when the time horizon changes.
- The creativity of creating good indicators.
Here’s some example of technicals indicators that can be used:
- The difference between the current price and its exponential moving average (EMA).
- The difference between two EMAs with different ranges.
- The difference between three EMAs with 3 different ranges.
- The difference between bought and sold volume representing order imbalance.
- The difference between order imbalance and its EMA.