There is nothing worse than misinformation, intentional or even unintentional, in finance world. Misinformation translates directly to money losses in financial markets. Here are some truths about non-stationary prices and machine learning in finance.
Tweets drive markets lately so let tweets drive this article too.
Beware of "gurus" driving #traders to exercises in futility (statistics with no end in sight.) #Trading and even long-term investing are possible because prices are non-stationary (trending.). The claim that non-stationarity is a problem for trading is beyond naive, maybe stupid
— Michael Harris (@mikeharrisNY) September 9, 2019
If prices were stationary, then markets would be too efficient to trade or profit from investments. The reason profits are possible is because trends form in all timeframes. Those that get the trends correct profit at the expense of those who get them wrong; this is a zero-sum game.
Therefore, trading is possible because prices are non-stationary. The objective should be to embrace non-stationarity and find ways to take advantage of it, not trying to find ways to eliminate it. Trends generate the bulk of the profits while picking bottoms and tops is usually an exercise in futility. Simple moving average crosses have dealt effectively so far in timing non-stationary price changes. For example, the trivial 50/200 long-only moving average cross has nearly matched SPY ETF annualized returns since inception at substantially lower risk, as shown below.
Anyone who displays obfuscated math to impress audience should first present a model that beats the risk-adjusted returns of above simple model. This is rarely done but just talk and formulas. The above model is just an example. There are many models that work well with non-stationary prices and realize returns. These models are non-stationary processes. Looking for stationary is beyond naive.
There is a lot of talk about fancy ML methods accompanied by obfuscated math but little proof they work. They may be good for academic curriculum but are of little value to most practitioners. There is also little talk about feature engineering. For example, our DLPAL LS software does the hard work of feature engineering. The ML models that use these features can be as complicated as the user deems appropriate but if features have little economic value (predictive value), then ML is GIGO.
@numerai has a good competition. If anyone thinks they have something of value in ML that applies to trading they can demonstrate it there. Unfortunately we cannot apply our features to their competition because their data are encrypted.
Two truths in conclusion:
- Non-stationarity is where the profits come from
- The key in ML is feature engineering not so much the fancy algos
If you found this article interesting, I invite you follow this blog via any of the methods below.
If you have any questions or comments, happy to connect on Twitter: @mikeharrisNY
Charting and backtesting program: Amibroker
Technical and quantitative analysis of Dow-30 stocks and 30 popular ETFs is included in our Weekly Premium Report. Market signals for longer-term traders are offered by our premium Market Signals service.