Also, should one use 2x ATR? 3x? Only way to have confidence is to run a backtest. And then re-run everytime markets condition change. SDs are an autobacktest all the way. Everybody knows what 95% of the distribution means, 3x ATR, what does that mean? I haven't a clue
Volatility (a particular measure of a standard deviation calculation) and Average True Range are both creatures of the very same price history, and are both equally responsive. If you want to prove this to yourself, tweak ATRs and Bollinger Bands for a bit.
If Stddev works visually, then that can be a plus. Likewise, ATR can be used for similar purposes, but will not work so well if you try to use it for curve-fitting, but rather as long distance stops based on price action potential (ie. long bars in consolidation). Statistically, none of these are "superior" for price series, as price series are non-stationary growth/decay-series, thus hard to predict and quantify. But, it really depends on usage of these tools and as all empirical studies, hypothesis' based on observations. You can even use EMA and WMA on true range (TR) in order to make them more responsive to recent price action. Multiplying works about the same, so will have similar "pseudo"-statistical properties. They're good to experiment with as another tool in the toolbox.
I think some of you guys have so much experience to be going to forum for this kind of answer, doesn't anyone back test? Surely you all have tick data over several years, find out from winning trades where you can keep approx 90% winners and see how much price went against trade. Find out what kind of time of staying in trade before it goes south. IMHO this is best ways for finding optimum stops. And if doing long term stocks/commodities, can use percentages.
Something I haven't seen discussed yet (although I have just scanned the answers), is what does hitting the stop do to your account, and how big is this stop in relation to your average win? If you're going to take CL as the example, and you put your stop in an area that price didn't go to 95% of the time in the last 60 days, does this mean the stop is like $5 away? This would mean that its a stop that costs you 5k when hit. If your trading only ever captures a 50 cent move, and hence $500 profit, then this stop is huge. If you hit this stop once or twice in the course of 20 or 30 trades, man oh man, that will be bad. So what is the stop in relation to the profit target? This is key.
Fair enough question, but the OP posed a σ versus ATR question. With either metric, "shit [can] happen" -- but it's more of a {worthy} Kelly/Tharp question... I have always relied on "42".....
Tried it early in my career. Your thinking is nothing new. It worked out for awhile and made a nice profit. Then it blew up. Stops are always tricky. I always found that there were a few trades that really messed up my P&L. The answer is simple, right, just go tighter with the stops. Once you got tighter, you stop out a lot of trades that end up winners.
That is a very interesting technique. Some questions on the stability of this tuned parameter: 1) How long before it goes out of sync? 2) Does it go out of sync within a predictable duration? The argument between ATR and SD is between the behavioral and mathematical view of the markets. I believe ATR has an imbedded psychological and implied volume component that most quants want to isolate from their market models.
Not sure what you mean by "out of sync" (whether σ out-of-sync with ATR, or whether σ & ATR out-of-sync with the underlying) but once the look-back is lined up, they will behave nearly identically. But for shorter duration, the ATR will tell you more clearly *what*just*happened* (so I favor it for weekly options). As to whether either σ or ATR might go out-of-sync "with a predictable duration" (again, not sure of the question, but...), if you're writing of immediate price excursions beyond the σ or ATR envelope ("envelope" being constant or Bollinger[ed] and dynamic) -- they would do it in similar fashion -- the handy thing about σ over ATR is that connection back to a Gaussian (or LogNormal or Pareto or Weibull or roll-your-own) Distribution, and the probability distribution you can infer, AND THE NUMBERS that follow: "since my [stop] is 1.64 σs out from current price, then if the extant distribution holds going back over time, I should be able to count no more than 5 instances of price outside of this _x_ value, over 100 observations." That's something that's just not so meaningful in the shorter term, but if you're talking 100 hours, or 100 days, your use of σ really gets you somewhere. (My opinion.) {The "holds over time" part lends use -- even *comfort* -- in using σ going forward, with the simple assumption that the past behavior ('variance' in this case) is representative of future behavior. This happy idea is represented in a bunch of the posts above. It also, though, shows a big fat *fiction* in so much of what we do: we depend on "I.I.D." -- independent and identical distributions -- meaning our dice cannot remember the previous outcome when we roll 'em again. "Ooops!" The market is *random*?!?!? Hardy-har-har-har. "But I digress......"}