Trend Followers Make Forecasts

There is an ongoing debate in some trading forums about whether trend-followers make forecasts. The debate started after in an interview, Michael Covel, who is an author, claimed that trend-followers do not make predictions. The debate intensified recently after Jim Simons trashed trend-following in an interview as irrelevant.

The best answer to this question that I know off was given in the blog. The author, who is a quant, called the claim that trend-following does not make predictions “complete nonsense”.  He used a simple forecasting model to show that all trading algos, including those that are used by trend-followers, do make forecasts.

Here we have to be careful. The term “prediction” can only be used in finance in the sense of forecasting and not in a supernatural sense. That would not make any sense anyway.

I approach this problem from the angle of logical reduction. I can think of three different types of trend-following models:

  1. Algo-based entries and algo-based exits
  2. Random entries and algo-based exits
  3. Random entries and random exits with a trend bias

I would think that the claim is that type (1) systems do not make forecasts because they ride an uptrend or downtrend and exit when some condition(s) are met. An example is the 50-200 moving average crossover system is SPY since inception, with an equity curve shown below:

Performance of long/short 50-200 moving average crossover system in SPY

Performance of long/short 50-200 moving average crossover system in SPY

Note how this trivial system has outperformed buy and hold at a lower max drawdown of -35%. CAR is 9.7% on a small sample of 20 trades and the win rate is 60%. Not bad for a trivial system that generated a short position in SPY at the open of last Friday. This is certainly a forecast of a downtrend. What else could it be?

A trend-follower could claim that the entries are essentially random and not forecasts are being made when they occur. This is the type (2) system. However, a forecast is made when the exit signal occurs and the position is reversed. Even if this was a long-only system, when an algo-based exit occurs, a forecast is made that the market will change direction. Therefore, type (2) systems also make forecasts.

Type (3) is a random system and far from one that could be used for trend-following. However, in SPY and due to the structural upward bias, 100% of random long-only systems with a random entry and exit sufficiently apart have generated gains. This can be shown by a simulation with the following properties:

  • A coin with p{heads} = 0.75
  • Toss a coin after the close of each day
  • Buy at the open of next day if heads show up
  • Exit long positions  if tails show up
  • Repeat the coin tossing until the end of price history
  • The simulation is repeated 20,000 times

The results of this simulation for SPY are shown below:

Simulation of random, long-only trading system in SPY with an uptrend bias

Simulation of random, long-only trading system in SPY with an uptrend bias

It may be seen from the distribution of the net return of the random systems that 100% of them profited. Actually, the lowest return was a little more than half the buy and hold total return of 566%.

But can we call this trend-following in the traditional sense? If trend-followers believe that they are random traders, then I agree that they can profit without making any forecasts in some specific markets due to their directional bias.

Actually, I use the three different types of systems with regard to their exits and entries in my new book, Fooled by Technical Analysis, to determine capitalization requirement for system trading.

The conclusion is that only random traders do not make forecasts. If some trend-followers think they are random traders, this is not too bad. In reality, 95% of traders do trade randomly

If you have any questions or comments, happy to connect on Twitter: @mikeharrisNY

Subscribe via RSS or Email, or follow us on Twitter.

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. Mean-reversion signals for short-term SPY traders are provided in our Mean Reversion report.

Copyright Notice

This entry was posted in Trend following and tagged , , . Bookmark the permalink.

8 Responses to Trend Followers Make Forecasts

  1. Jake Yeager says:

    Hey Mike,

    Perhaps part of the issue in this discussion is the definition of "forecast." If by "forecast" one means a definitive price that will arise sometime in the future, then trend-followers do not forecast. On the other hand, if you define "forecast" as a probabilistic determination of the market moving in a particular direction, although the extent of the move is unknown and not included in model calculations, then trend-followers do forecast.

  2. James says:

    Isn't this just semantics?
    All trading is based on forecasting: Use a method that gets high-probability outcomes either through high win-rate or with avg win/avg loss >1. Trend followers tend to use the latter. Whether or not trend following works is a matter entirely separate to 'forecasting' or 'prediction'. Every trader using a non-random system is a 'forecaster' – a forecast that they'll make money using their system. Swap the term 'forecast' for 'predict' and it makes no difference.

    I believe the debate to which you originally referred was about the validity of predictive vs. reactive TA – points you have clarified well already. Trend followers don't make forecasts or predict, meaning that they don't use predictive TA.

    BTW – I dont think Covel was a legendary turtle, but an author who collated and wrote about the turtles' methodology.

    • Hello James. Yes, you are right, Cover is a writer about TF. I made the correction. Thanks. As far as predictive vs. reactive TA, I think the latter is only used to describe the state of the market. You are correct here: "Every trader using a non-random system is a 'forecaster' " That was my point too.


    • tetrapharmakon says:

      I almost jumped on my seat when I read the excerpt from your article – since then corrected, thank you for that – from the Quantocracy mailing list that M. Covel was a turtle (or even a trader for that matter). Much of the recent angst relating to predicting vs reacting is generated and fueled by his ability to pitch CTA funds and their methods, which is the only reason I can think of why such a lot of legendary CTA manager (I mean, some serious, track-recorded, real traders) crowd to speak on his show on a regular basis. He might not be a trader, but he sure knows how to promote an idea.

      He often states the "non predictive" part is a feature of (unique to) *vanilla* (turtle inspired, mostly futures) trend following, which is plainly false. On the other hand, his show is targeted towards laymen who don't know the difference between a systematic approach and snakeoil newsletter advice and / or news anchors, the latest being in the "paranormal" sector. I think this is the point of his ranting on the subject.

      Now, as you clearly illustrate in this article, of course any trading system is predictive. An edge is a repeatable (statistical) advantage on the market. If you can't have faith in your edge you plainly don't trade. One trades because on predicts / forecasts that in the future his system will gain more than it will lose and without the account's equity visiting ground zero in the meantime. It is by its essence a form of prediction.

      Momentum being probably one of the most resilient market anomalies / inefficiencies, it has not been arbitraged away (yet) on most commodities, and a few other asset classes depending on the timeframe. Its worked for hundreds of years (many studies, most recent like Greyserman and Kaminski's) and we can *assume* it will continue to do so in the future. But this is an assumption, and in itself another layer of forecasting. Though more subtle and intrinsically superior to a news anchor's morning speech.

      So in the end, my position is that its a semantic problem.

      Thanks for the read 🙂

      • Thanks for the comment.

        You should watch this interview of Jim Simons:

        Simons claims that trend-following ended in the 1980s. I also claim the momentum died in 2000:

        The systems that some claim do trend-following (price series momentum) or relative momentum are not doing exactly that but something else. Their users actually do not know that. I am working on a paper about this.

        Anyone who claims that there can be a system or method that is not random and can make systematic profits without making predictions is either naive or a crook.


  3. Bo says:

    When talking about "forecasts", my first impression is what we often heard on CNBC when they interviewed some people like say Marc Farber who would say something like: the US stock market valuation is all time high so it is like it would crash 30% in October etc etc. This is a forecast and trend following does not make any forecasts in this sense. Some of the quant models do make such kind of forecasts.

Leave a Reply