I have been playing around with some different active trading strategies by looking at historical stock data. However, I am running into some difficulties in how to define whether a strategy actually "works".

For me, the key to a successful active trading strategy is to have it achieve greater returns than simply "buying and holding" the stock. If buying and holding results in better returns, then the effort of active investing is useless.

The Problem: Looking at historical data, I don't know where to stop and start the evaluation of my returns compared to buying and holding.

For example, below shows a historical chart where yellow dots are "buys" and red dots are "sells" based on some method.

Yellow dots are "buys" and red dots are "sells"

From the first buy to the last sell, the stock makes $26.45. Using my active strategy, I can make $28.23 during that same time period, thus out-performing the passive buy-and-hold (note that I neglect any trade commission charges here).

However, if I choose to evaluate my method by looking at the full chart, then the active strategy fails miserably. Buy-and-hold results in gains of nearly $55 between 2013 and 2019 while the active method only gets the $28.23 because no buy/sells are triggered using the method before mid-2014 or after mid-2018.

If I had used this method and only looked at the time period from mid-2014 to mid-2018, I would've walked away thinking that the method is pretty good. But if I looked over a broader time range from 2013 to 2019, I would conclude that the active strategy was abysmal. Similarly, if I re-do this analysis in mid-2020, I might come to some other conclusion depending on how the stock behaves going forward.

The choice of timeframe to evaluate returns ultimately seems arbitrary. Is there some guideline to follow?

3 Answers 3


Trending indicators work well in strong trending markets. They're late in and late out. They're not effective in sideways markets because they will generate many false signals and whipsaws.

Oscillating indicators generate more signals and they are best used in trading markets. However, more signals means more false signals and whipsaws.

The problem isn't a function of knowing 'where to stop and start the evaluation of returns compared to buying and holding.' It's a problem due to return being dependent on the appropriate type of indicator being used at the appropriate time. And how does one know what that was? Hindsight. There's no way to know what will work best, going forward.

Here's another way to observe this. Optimize a trading system on the first half of your security's historical data. Then test it on the second half. Odds are, it won't perform well. Test it on other sets of data. For some, it will work splendidly. For others, abysmally. At some point, you'll realize that using this isn't the road to riches.


Imagine two scenarios:

  1. Active trading as you have above.

  2. Buying and holding over the same period of time.

The natural comparison to draw is between how much you made over the whole time-frame, from your first buy to your last sell because that's what you would have with your buy and hold strategy. Buying and holding is the null hypothesis, you are trying to see whether your strategy of actively buying and selling is able to beat it, if it's not able to beat it over the whole time period then what's the point of doing it? You'll end up spending more time and energy on making those buys and sells and analyzing market data, and pay more in trading fees. So what are you getting from it?

The only potential upside that I could see to active trading as you are doing it here is if your strategy works well over a very long period of time, especially during a market crash. If you could show (by crunching historical data) that your strategy outperforms the market when things get bearish, and you suspect an imminent bear market then it might make sense to continue using your active strategy. The only issue with this approach is that your model was made using historical data so it may be very good at recognizing the signs of crashes that have previously occurred but not ones that will occur in the future. One way to get around this would be to test it on a different similar historical data set -- if you're analyzing large cap stocks how does it work on medium cap? What about other markets? What about similar companies with longer histories that overlap financial crises?

  • You made some valid points. Some additional thoughts. Currently, you won't pay more in trading fees since multiple discount brokers have recently eliminated commissions. The trading strategy in the posted graph would have received interest on cash balances, since there were long periods when out of the market (unknown if the OP accounted for that). What would be significant to me would be if the trading strategy avoided huge downdrafts such as in 2000 and 2008. I'd take that and identical performance over B&H in a heartbeat and that would be the point of doing it. Oct 18, 2019 at 16:34

Using spectral analysis, the periodicity of the slope estimators is 40-41 years. Yes, not joking. That is the amount of time it appears to take evaluate any strategy from a formal statistical sense. I know this because I was curious for a practical reason I performed spectral analysis by ranks and on the logarithmicly transformed data.

A simple way to understand spectral analysis would be how long it takes to see everything once. An example of that would be the weather and a year. If you made your strategy in December and implemented it in July, you may be in for a rude surprise.

Sorry, I no longer have the technical results. Because most firms do not last 40-41 years, you will need to use information sharing via a Bayesian prior. Except for a highly censored set of long-lived firms, you are not going to get an unbiased estimator that has any validity for a shorter period.

Since it appears you can code, you can test this yourself.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .