Skip to content
Get App

I Tested 8,080 Trend-Following Strategies. Only 10 Survived.

Algovibes tests 8,080 trend-following variants across 20 assets; only 10 survive statistical correction, all Bitcoin with volatility-scaled sizing.

Key Takeaways

  • Testing thousands of rules guarantees some lucky winners; statistical correction is essential before claiming an edge.
  • Only 10 of 8,080 configurations survived correction, all Bitcoin with volatility-scaled sizing, a very narrow finding.
  • Volatility-scaled sizing meaningfully reduces drawdowns but cannot fix every blowup, as the Exxon case shows.
  • At the diversified portfolio level, trend following reduces drawdowns but underperforms buy-and-hold on risk-adjusted and absolute returns.

What the video covers

  • The video tests 8,080 trend-following configurations across six rule families, five asset classes (stocks, commodities, indices, crypto, bonds), long-only and long-short, with fixed and volatility-scaled sizing.
  • A coin-flip analogy illustrates why many trials produce lucky-looking winners, motivating a statistical correction for multiple testing.
  • About 19.5% (1,576) beat buy-and-hold out of sample, but only 10 (~0.12%) survive the full statistical correction.
  • All 10 survivors are Bitcoin strategies, and none worked with fixed-size bets; all required volatility-scaled position sizing.
  • Volatility scaling sharply reduces worst drawdowns for many stocks (e.g., Apple from 99.98% to ~90%) but not for Exxon, where it barely helps.
  • At portfolio level, trend following loses on risk-adjusted return (73-83 vs. buy-and-hold 87) but cuts max drawdown from ~48% to under 13% with volatility scaling.
  • Buy-and-hold returns ~16% annually versus roughly 3-5% for trend portfolios, since avoiding crashes also means missing recoveries.
  • Trading-cost sensitivity shows a gentle decline in risk-adjusted return, but trend never beat buy-and-hold to begin with, so costs are not the decisive factor.
  • Limitations include survivorship bias (only currently listed tickers), a short test period, and the need for out-of-sample overfitting checks.
  • The video explicitly warns that the findings are narrow and should not be read as proof that trend following works.

Answers

Questions about this video

How many trend-following configurations survived the statistical correction?

Out of 8,080 tests, 1,576 beat buy-and-hold out of sample, but only 10 survived the full statistical correction for multiple testing. All 10 were Bitcoin strategies using volatility-scaled sizing.

Does the video conclude that trend following works?

No. The presenter explicitly says the narrow Bitcoin result does not mean trend following works in general. At the diversified portfolio level, trend following loses to buy-and-hold on risk-adjusted return, though it substantially reduces drawdowns.

Does volatility scaling always prevent catastrophic losses?

Not always. It meaningfully reduced drawdowns for most stocks, such as Apple from 99.98% to about 90%, but for Exxon the improvement was minimal, with the scaled version still losing 99.83%.

Full Transcript — Download SRT & Markdown

00:00
Speaker A
Trend following is sold as one of the few systematic strategies that keeps almost surviving rigorous testing.
00:07
Speaker A
Emphasis on almost. Most people testing it use one or two textbook signals. Price above its moving average may be basic momentum. Run it once and call it a day. I did not call it a day. I tested six rule families
00:24
Speaker A
across five asset classes with two different ways of sizing the bet for a total of 8,080 separate tests.
00:34
Speaker A
Same honest question everyone else is asking just answered properly for once. That's 20 instruments across five asset classes. Stocks, commodities, equity indices, crypto, and bonds. and bonds are in here because you can't call something crosset and then quietly leave
00:52
Speaker A
out interest rates. That's not cross asset. That's just the assets you already had a chart for. The six rule families cover price versus its own average plain momentum channel breakouts based on the classic turtle style approach, dual moving average
01:09
Speaker A
crossovers, a filter that only trades when a separate indicator agrees the market is actually trending, and a trailing stop system that tries to ride a position instead of flipping the second the signal twitches.
01:23
Speaker A
Each of those gets tested long only and long short, both with a plain fixedsiz bet and with position sizing that scales down when markets get more volatile.
01:34
Speaker A
Multiply everything out and you end up with 8,080 tests. I did not enjoy running all of them. Now, here's the part people usually skip, and it's probably the most important part of this entire video, so bear with me for 30
01:50
Speaker A
seconds. Imagine 8,080 people flipping a coin. Just a normal coin with absolutely no skill involved. Some of them are going to land five heads in a row. And a few might even land 10. Not because they're unusually talented coin
02:07
Speaker A
flippers, but because once you have 8,000 people flipping, somebody's luck is eventually going to look like talent.
02:15
Speaker A
Now, imagine I hide everybody else and only show you the winner. 10 heads in a row. Incredible system. You'd probably want to know what happened to the other 8,079 people.
02:29
Speaker A
And that's exactly the problem with testing 8,080 trading rules. Some of them are going to beat buy and hold by accident. Not because the rule actually works, but because I tried enough rules that a few impressive looking flukes
02:43
Speaker A
were basically guaranteed to show up. So before I let anything call itself convincing, I ask the annoying question, if none of these rules actually had an edge, how many lucky winners would I expect to see anyway just from trying
02:57
Speaker A
this many? And then I ask whether what I actually found clears that bar or whether it's sitting exactly where dumb luck would have put it. That's what the statistical correction is doing here. It doesn't make anything look better. It
03:11
Speaker A
just refuses to get excited about a result until that result beats what coincidence alone could plausibly hand me for free. So, here's what survived.
03:20
Speaker A
We start with 8,080 tests in total. 1,576 of them beat their own buy and hold benchmark out of sample. That already sounds pretty good until you correct for the fact that we tried 8,080 different things.
03:38
Speaker A
Once you ask which of those apparent winners actually clear the statistical luck bar, we're left with 10.
03:46
Speaker A
10 out of 8,080. The other 1,566 looked interesting, but they didn't survive the correction for how many chances we gave ourselves to get lucky.
03:58
Speaker A
So roughly 19.5% beat buy and hold out of sample, while only about 0.12% survived the full statistical correction.
04:08
Speaker A
That's not a particularly impressive hit rate for a strategy family with its own section in practically every trading book ever written.
04:17
Speaker A
But here's where it gets even more interesting. All 10 survivors are Bitcoin strategies. Not crypto broadly and not Bitcoin plus Ethereum. Zero Ethereum variance made the cut. Across this entire test universe of stocks, bonds, commodities, indices, and crypto, every configuration
04:39
Speaker A
that survived the statistical correction came from Bitcoin. And there's a second catch. None of those 10 survivors worked with a plain fixed size bet.
04:50
Speaker A
Every single one required volatility scaled position sizing, meaning the strategy takes a smaller position when markets get wild and a larger one when volatility comes back down. Use the same fixed bet regardless of what the market is doing. And even those Bitcoin results
05:07
Speaker A
stop clearing the statistical bar. And before anyone clips that sentence out of context, this does not mean trend following works. It means that within this particular data set, this historical window and this specific testing framework, a very narrow group
05:24
Speaker A
of Bitcoin strategies survived the correction and only when volatility scaled sizing was used. That's about as narrow a finding as you can have and still call it a finding. So, it should be framed accordingly.
05:39
Speaker A
That sizing result raises another question, though. is using the same fixed position regardless of market conditions. The reason some of the individual stock strategies end in what I believe is technically called a bloodbath.
05:55
Speaker A
So I tested that directly. Same rules, same instruments, once with a flat bet and once with volatility scaled sizing.
06:03
Speaker A
Apple trading both directions without volatility scaling reaches a worst case draw down of 99.98%.
06:12
Speaker A
So essentially your entire account is gone. Scale the position by volatility instead and the draw down falls to around 90%.
06:22
Speaker A
Still catastrophic obviously but genuinely better and broadly the same pattern appears with JP Morgan, Caterpillar, Home Depot and most of the other stocks.
06:34
Speaker A
Exxon however did not get the memo. Its unscaled draw down is 99.9986% and the volatility scaled version still loses 99.83%.
06:48
Speaker A
That's less of an improvement and more like rounding error with extra steps. So volatility scaling definitely helps sometimes by a lot but it does not perform miracles.
07:00
Speaker A
Now we get to the part where diversification is supposed to save the day. Instead of looking at one instrument at a time, I spread the same trend following approach across all 20 markets and look at the portfolio as a
07:15
Speaker A
whole. And the headline number here is the honest average across all 101 tested variants, not whichever individual setup happened to produce the prettiest equity curve. Because picking the best looking configuration after the fact and calling that the finding is exactly the kind of
07:34
Speaker A
thing this entire project exists to catch people doing including me. The actual numbers are pretty clear.
07:43
Speaker A
Trend following produces a risk adjusted return of 73 with a flat bet and 83 with volatility scaled sizing. Buy and hold comes in at 87.
07:55
Speaker A
So trend loses on riskadjusted return in both versions. But the drawdown story is genuinely impressive. Buy and holds worst stretch leaves the portfolio almost 48% underwater while trend with a flat bet cuts that to around 22%.
08:14
Speaker A
With volatility scaled sizing the worst draw down drops below 13% less than a third of buy and holds worst crash.
08:24
Speaker A
That's a real improvement, but you pay for it. Buy and hold makes roughly 16% a year over this test, while the trend portfolios make somewhere between 3 and 12 and 5%.
08:37
Speaker A
Avoiding the worst parts of a crash also means you can miss part of the recovery afterward. And historically, that recovery is where a lot of the return gets made.
08:50
Speaker A
The next obvious question is whether this result is just one aggressive trading cost assumption away from disappearing completely.
08:59
Speaker A
So I ran it again at half the assumed trading cost at the baseline cost and at double the cost. The risk adjusted return moves from 76 to 73 to 66.
09:12
Speaker A
That's a gentle slide rather than a cliff. But there's an important point here. trend following didn't beat buy and hold on riskadjusted return in the first place. So saying the edge might disappear after costs isn't really the right conclusion. At this level of
09:29
Speaker A
aggregation, there isn't an excess risk adjusted edge here for the cost to destroy. And then we get to the last test, the one people tend to skip because it's usually the least flattering.
09:42
Speaker A
Overfitting. If I use historical training data to pick the configuration that looks best, how much of that apparent advantage survives when the chosen configuration finally reaches data it has never seen?
09:56
Speaker A
I tested that over four separate rolling windows, selecting the best configuration from roughly 400 possibilities each time, then freezing that exact choice before evaluating it on the future period.
10:10
Speaker A
On the data used to select the strategy, the average sharp is 0.90. On the unseen data, it falls to 042.
10:20
Speaker A
So more than half of the apparent risk adjusted performance disappears once the strategy leaves the data it was selected on. And roughly one out of every six times, the configuration that looked best during selection actually goes negative when it reaches the unseen
10:37
Speaker A
period, which is statistically speaking not great. There are also a few limitations worth being completely open about before wrapping this up. Every ticker in this test is currently listed and currently liquid, which means survivorship bias is still present and flatters both trend
10:57
Speaker A
following and buy and hold. The trailing stop design is one reasonable implementation among several. So changing that design would move the numbers for that rule family. And when I calculate the average across the 101 portfolio variance, I'm treating every
11:14
Speaker A
rule and every parameter setting as equally important. I think that's defensible, but it's still a modeling choice. And a different waiting scheme would produce a different number that I haven't tested here. So after all of that, we can reduce the entire
11:31
Speaker A
experiment to five questions and five fairly straightforward answers. First, does diversified trend following beat buy and hold when you test it properly?
11:43
Speaker A
No. It loses on risk adjusted return in both sizing versions. And doubling or h havinging the trading costs doesn't change that conclusion.
11:54
Speaker A
Second, does trend following still reduce draw downs? Yes, substantially. The worst draw down falls by more than half with a flat bet and by almost three quarters with volatility scaled sizing.
12:08
Speaker A
Third, does any individual strategy survive serious statistical scrutiny once we account for the sheer number of tests? 10 out of 8,080.
12:20
Speaker A
Everyone is Bitcoin and everyone uses volatility scaled sizing. None survives with a plain fixedsiz bet.
12:28
Speaker A
Fourth, does volatility scaled sizing fix the worst blowups? Mostly, Exxon didn't get the message.
12:37
Speaker A
And fifth, is the overfitting risk real? Absolutely. Average sharp falls from 0.90 during selection to 42 on unseen data. And around 1 in six of the configurations selected as the best actually turn negative when they reach the period that matters.
12:56
Speaker A
None of those answers contradict each other. And that's really the whole point of testing something properly instead of running one back test and declaring victory.
13:07
Speaker A
8,080 tests didn't hand me a magic signal to trade tomorrow morning. What they gave me was a much more honest picture of what trend following seems to do well, where it struggles, and how narrow the strongest evidence actually
13:21
Speaker A
is. That's the honest result. If you want the full picture, the complete code behind this study, every other strategy I've tested, the data sets, and the live experiments, it's all inside the Algo Vibes research lab. And I don't mean
13:38
Speaker A
just the highlights. The winners are there, the failures are there, and you get the tools to inspect the results and check the claims yourself. Members also vote on what gets tested next. No signals, no promises, just the research.
13:55
Speaker A
It's atlab.wisequant.app. Thanks a lot for watching and I will see you in the next video.
Topics:trend followingbacktestingoverfittingvolatility scalingBitcoinstatistical correctiondrawdownbuy and holdquantitative tradingalgorithmic trading

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries and speaker detection. 30 minutes free.

Or transcribe another YouTube video here →