Invest Without Emotions
Open sourced by Meta back in 2017, Prophet is a procedure for forecasting time series data. Here, we give it two years of monthly returns and use it to forecast returns one month forward. The forecasts are ranked and a portfolio is constructed for the forward month.
The results are not as bad as fitting a simple linear model but not that great either.
Code on github.
More often than not, simple models outperform complicated ones. Inspired by some recent academic research that showed that linear regressions yielded better momentum performance, we did a quick backtest to check if building a linear model through recent 12 and 1/3/6-month performance and creating a portfolio using its next-month predictions made sense.
Counter-intuitively, a naïve momentum strategy outperformed linear models.
This is not our first run-in with linear regressions. Our Dynamic Linear Model strategy simply regresses prices to a 45* line and ranks them based on goodness of fit.
Most of the time, of all the different ways to skin the cat, the simplest is the best one.
Code on github.
Our previous posts introduced portfolios based on Moving Average Distance (Part 1, Part 2). To answer questions regarding the stability of the moving average lookbacks, we ran a rolling window, picked the “best” MA lookbacks and walked the portfolio forward by a month. We expanded the window through 12 to 60 months in 12 month increments.
Turns out, most of them fall within the 20/200 region.
The data-mined parameters create portfolios that perform on par with the 21/200 used in the paper. While we are always skeptical about magical parameters that make the research work, at least in this case, the magic is not too far fetched.
You can follow along the live version of the original strategy here: MAD 21/200
Our previous post introduced a paper that used a moving average crossover to create a portfolio of stocks. While the backtest using the parameters in the paper looks good, the presence of these “magic” lookback parameters gives us pause. Did the authors just try a bunch of different parameters and published what worked? What if we do an exhaustive search through all possible combinations?
Here are the annualized returns and Sharpe ratios pre-COVID:
The magic 21/200 lookbacks look legit. However, the post-COVID picture looks different:
The magic parameters don’t quite figure in the top 5. However, even if you used the data-mined set, you would be ok?
Also, the paper used a “sigma” parameter as a threshold to activate the crossover. Getting rid of it seemed to have lopped 10% off the post-COVID returns.
You can follow along the live version of the original strategy here: MAD 21/200
Code and charts on github.
Sometimes, a research paper comes along that gives academic rigor to an obvious thing that trend-followers were doing for decades and makes you sit up and take notice. Moving Average Distance as a Predictor of Equity Returns, Avramov, Kaplanski and Subrahmanyam (SSRN) does just that.
Turns out, a simple moving average crossover signal proves robust to momentum, 52-week highs, profitability, and other prominent anomalies.
A later paper extends it to international stocks and finds similar results (SSRN).
A quick backtest shows that it works for Indian stocks as well.
It looks like COVID turbo-charged this strategy. The pre-COVID equity curve is saner.
The returns are good but it comes with some nasty drawdowns. Not sure if most investors can stomach a 25% drawdown that lasts over a year. Can it be made better by applying a volatility filter?
By sacrificing 2 points of returns, you can get to a sub 20% drawdown. Also, the filter worked during the most recent 2021-23 drawdown as well.
You can follow along the live version of this strategy here: MAD 21/200
Code and charts on github.