Trading "Noise"

The Success of Simple Strategies for Retail Trading

The recent rise in retail trading has drawn lots of media attention about potential market instability and manipulation. While Reddit and Discord are now common places for retail traders to exchange trading ideas and publish their own due diligence about certain stocks, many new investors are left asking themselves: Are there simple strategies, which retail investors can use to consistently outperform the market, or are all active retail investors doomed to ultimately end up with a Sharpe Ratio of -1? Can retail traders systematically trade and earn excess returns, or will the retail trader inevitably become a noise trader?

In finance, noise typically refers to information or activity that confuses or misrepresents genuine underlying trends. Noise tends to be regarded as a problem for traders and investors. When there is lots of noise, traders cannot identify trading signals in the underlying, which leads to poor performance of their strategies compared to the market (Sharpe Ratio below 1). Hence, the negative term noise trader refers to a trader who confuses noise for signals. However, there are ways in which noise, if interpreted as a random deviation from a trend, can be traded to the benefit of the trader. In this post, I show that small-scale traders can trade noise with a very simple strategy and reap significant risk-adjusted excess returns.

Success of a Simple Strategy

In fact, the strategy I use is so simple that it requires no asset-specific research or active monitoring and only trades US stocks during less than a one-hour time window. I will not provide further details about the logic of the strategy (for obvious reasons). I forward-tested the strategy since January 2021 in a paper account with IBKR using their API with a simple trading bot in Python. The strategy outperformed S&P500, QQQ, as well as any individual Fama-French factor significantly and is on track for a yearly Sharpe Ratio of 3. Below you can see the results of the strategy’s forward test for quarters 1 and 2 of 2021 compared to the performance of QQQ and the S&P500.

The forward test in a paper account is more realistic than any backtest, because it accounts for fees, spreads and execution quality. Only some doubts remain: Does the paper account performance translate into an equally good live performance? What about the tax impact? While I can say nothing about the former concern, tax is no major issue for this strategy, as it does not commonly repurchase assets within a month after selling at a loss. It is worth noting that this noise trading strategy is not necessarily reproducible for large-scale traders. It relies on a quick exploitation of asset- and time-specific pockets of opportunity, which may not persist long enough to realise large trades. However, any retail trader (of say up to 1m invested volume) would be able to use a strategy of this type.

Origin of \(\alpha\)

VariableEstimateP-value
\(\alpha\)0.56%1.69%
QQQ-2.660.16%
S&P5003.5515.44%
MKT-0.2192.24%
HML-1.360.72%
RMW-1.0910.22%
CMA1.401.54%

Should I use Machine Learning?

Machine learning can extract information from data, right? So it should be useful to create trading strategies? This type of thinking has led many young people with an interest in financial markets and education in machine learning methods to consider searching for trading signals in financial data with machine learning methods. While some people may have success with this type of workflow, in my opinion it is suboptimal at best.

The large amount of noise in financial data complicates signal extraction. Unlike in engineering applications, where there is often a clear signal, machine learning methods face significantly tougher circumstances in financial data. In simple terms, it is much easier to identify a dog in an image, which any human is also capable of, than to know whether e.g. an ascending triangle on a 15min chart will lead to the positive breakout a technical trader would typically expect. However, even a small improvement of the predictive R-squared for excess returns can imply significant positive returns. Gu, Kelly and Xiu (2018) provide a comprehensive summary of (successful) machine learning applications to asset risk premia forecasting. While neural networks and tree-based methods provide economically significant returns (Sharpe Ratio above 2), they also find a relatively small set of most successful common predictors. Those include price trends, volatility and liquidity.

It is certainly expected that price trends, volatility and liquidity determine asset risk premia. Many investment strategies are partly based on price trends, like momentum and mean reversion. Due to risk-aversion in investors, high-volatility stocks will need to pay a high risk premium, which creates a natural link between asset risk premia and volatility. Trading illiquid stocks (e.g. microcaps) usually involves large bid-ask spreads, fees and execution difficulties (slippage), such that apparent price inefficiencies cannot be exploited by traders. Hence, it remains uncertain how Gu, Kelly and Xiu (2018)’s results are novel or useful for traders and investors. As their research only identified well-known determinants of risk premia as important predictors, retail investors should ensure that they understand price trends, volatility and liquitidy concerns first. Note also that the best possible Sharpe Ratio obtained by Gu, Kelly and Xiu (2018) is 1.69, whereas the much simpler noise trading strategy above is on track to realise a yearly Sharpe Ratio of 3.

Conclusion

While machine learning can help traders and investors, a deep understanding of market psychology and noise remains the most useful sources of inspiration for successful retail trading strategies. Pockets of opportunity can result from small market inefficiencies, which often cannot be exploited by large-scale traders and investors. However, retail traders can utilise those small inefficiencies, which are often interpreted as “noise”. Retail investors can reap significant risk-adjusted excess returns with very simple strategies. Machine learning methods, as fascinating as they are, are most useful when applied to an already successful trading strategy.