Stop Loss is not always your friend...

[KnowThePast]

Re: Stop Loss is not always your friend....

Apologies for being late to the party, I wanted to give this some proper research and time was short over the last week.

My intention was to run a few backtest that clearly show that stop losses are not your friend in a fundamentally oriented, long term portfolio. A little bit of bias on my side. I remember running some tests a long time ago that didn't inspire me, I thought I'll do it properly this time.

To start off with, I establish a benchmark - which is a return I would get by investing into all stocks in my universe. I set a period of 5 years, and it happened to give me a return of 68.41%.

Then, I select a strategy that beats the benchmark. Let's use the old and tried Net Assets. We buy when P/B < 0.7, we sell when P/B > 2. Position Size = 2%, brokerage = $30. Here's how it compares to the benchmark:



Higher return and more winners, but also more big losers. Nothing controversial so far. Let's now add stop losses to this winning strategy. I'll backtest on stop losses at 5%, 10% and 20%. Once stopped, share will not be bought again the next 365 days. Here are the results:



Equity Curve:


Hold on, that's not right, not the result I was expecting. Perhaps it just happened to work with this strategy, let's try a low PE strategy. Equity curve:



An identical result to the P/B strategy - stop losses improve the result, and the tighter the stop loss the better.

How about trailing stop losses? Not so good:


I am limited to 5 attachments, to be continued in next post.

 

[KnowThePast]

Re: Stop Loss is not always your friend....

Then it dawned on me - I was assuming that a 5% stop loss would execute at a price exactly 5% (rounded to nearest $0.01) below purchase price, which is not going to be the case. So, I've put in 5% slippage, each 5% stop loss would actually execute at 10%, 10% at 15%, etc. This brought the result down, but still it is substantially above the benchmark:



I tried a High PE strategy, that underperforms the benchmark. I'll spare everyone looking at another graph - the result, again, was that stop losses improved the performance of even an underperforming strategy.

This analysis stands at complete odds with RY's statement that stop losses do not improve the expected return and are simply a risk management mechanism, which is likely to bring down returns, not improve them. When my analysis disagrees with RY, my default stance is that I made a mistake. What did I miss?

RY, how was your data backtested? I ask, because I know standard backtests that re-balance yearly may not be very suitable for testing stop losses.

I ran a P/B strategy for the last 6 months, with a 5% stop loss and 5% slippage. That beat the same strategy without stop loss by a whopping 16.13%. Attached is a spreadsheet with all the trades, in case anyone is kind enough to look at it and spot an error.

View attachment stoplossTrades.xlsx

 

[DeepState]

Re: Stop Loss is not always your friend....

This is the definition of a positive expectancy. Lots of people talk about positive expectancy and it is easy to understand as an 'outcome' but what produces a positive expectancy? An edge I hear people say - but what is an edge? Well the normal answer seems to be anything that gives a positive expectancy, which just puts us into a loop, where the question of what is an edge doesn't need to be answered.

Do we need to understand what an edge is if we can observe its positive expectancy?

If you don’t understand it how do you know if it has stopped working until after you observe its outcome as a negative expectancy and the damage is done to your account?

If you are relying on edge identification purely through historical expectancy outcome, How accurate is your data? is it a valid positive expectancy or invalid data.


Even if your data is perfect, how do you know it is a robust edge and not a data mined expectancy? Huge numbers of variables you can dream up will have positive expectancy on historical data just through randomness without any likelihood of future utility.


Is it valid to only identify an edge as positive expectancy outcome?

If not, what than is an edge?



Are you saying the only possible edge is accurate prediction of price?

....

What's this got to do with stops?

IMO you cant define if stops are a risk/money management expense or an integral part of the process that creates your edge until you can define the 'cause' of your edge.

An edge generates positive expectancy. For this purpose, it doesn't matter what the edge is. It is just something whose predictive power in relation to a price, over some time frame, is above zero.

You do not have to understand it. There is no law of investment that says you need to. However, not understanding it increases the chances that you are just using a data-mined outcome and thinking it has edge when it is actually static that has produced the outcome. Expectancy is a concept of tendency. Outcome is a draw from the distribution of possibilities around that tendency. For every real predictive driver, there are many many more non predictive ones that can produce an outcome at least as good over some period by chance.

If you have a good notion of what drives the positive expectancy, you have a chance to know when the driver is broken. For example, the dividend run-up phenomenon is stronger for franked stocks than not. The reasons are obvious. If the tax laws change, you do not need to wait 10 years for a statistically significant sample to be created. You cut it on the announcement of the new legislation, before the next data point is collected, and move on.

You can derive a judgment on expectancy any way you like. A lot of methods, including movement of hands, is visible just by scoping around the threads. However, there are a great many ways to narrow in on whether what you are seeing is real or a statistical anomaly. The people who actually do this for a living will observe the economic transmission channels through which the result is found to check for validity. They will check it through time, around the world, through different regimes, through different types of stocks (value/growth/large/small/industries/foreign exposed/volatile...) even seeing how it might flow through balance sheets...as an example. It gets much deeper, but you get the idea. Basically, a statistical outcome is no-where near sufficient to say you have positive expectancy. About 95% of the things that seemed interesting on first round are dismissed on further analysis, in my experience. The stat is just a data point. It is reasonable to use something even if the positive expectancy is not shown in the statistical data.

You raise a good point about the definition of edge. I defined it for myself and indicated that other variants are also valid. Breaking things into return and risk is pretty clean, so I use this framework for such purposes. This does not mean that this is the only framework. For example, you can argue that an edge is anything that improves risk adjusted return or utility or liability mismatch. These combine notions of return and risk. However, given this chain of conversations related to asking whether stops make money in and of themselves, the cleanest framework seemed to be return and risk separately.

Stops do not make money. An edge that creates genuine positive expectancy is the only thing which generates returns on average, through time. Stops do not make money on average through time and reduce your expected profit if you have positive expectancy. In trying to explain this, it is important to separate what makes money and what does not. Edge makes money on average, through time. You will see evidence of belief that stops create expected returns throughout the threads in and of themselves. It cannot.

Risk management is important. I am not doing away with it at all. However, risk management does not make money. It determines magnitude of risk and can alter the shape of your outcomes. You can reshape the distribution via stops (let's not get into mispriced options) to change hit rate, for example. Placing trailing stops will increase hit rate. But it will change the reward profile to offset it in a way that will not change the central expectation. To suggest otherwise is to move into alchemy.

You are almost never sure of what your edge is. Position sizing should reflect your uncertainty. That is different to saying position sizing makes you money. It helps produce a superior risk adjusted outcome.

All this pedantry is required here just to divide the role of stops into risk management and/or return generation. It sits squarely in risk management. It has a valuable role to play in regard. I use stops, for example.

I should add, an edge tends to be defined in contexts related to alpha and (sub) zero-sum situations.

 

[DeepState]

Re: Stop Loss is not always your friend....

Then it dawned on me - I was assuming that a 5% stop loss would execute at a price exactly 5% (rounded to nearest $0.01) below purchase price, which is not going to be the case. So, I've put in 5% slippage, each 5% stop loss would actually execute at 10%, 10% at 15%, etc. This brought the result down, but still it is substantially above the benchmark:



I tried a High PE strategy, that underperforms the benchmark. I'll spare everyone looking at another graph - the result, again, was that stop losses improved the performance of even an underperforming strategy.

This analysis stands at complete odds with RY's statement that stop losses do not improve the expected return and are simply a risk management mechanism, which is likely to bring down returns, not improve them. When my analysis disagrees with RY, my default stance is that I made a mistake. What did I miss?

RY, how was your data backtested? I ask, because I know standard backtests that re-balance yearly may not be very suitable for testing stop losses.

I ran a P/B strategy for the last 6 months, with a 5% stop loss and 5% slippage. That beat the same strategy without stop loss by a whopping 16.13%. Attached is a spreadsheet with all the trades, in case anyone is kind enough to look at it and spot an error.

View attachment 60698

....

The result comes out of options theory. It's fairly straight forward. Given your quant skills and for those inclined to stochastic thinking, here's the argument in a highly simplified framework that actually goes to the binomial lattice methods for options pricing.

• You have a fair coin (ie. no forecasting power).
• We flip this coin 10 times. Heads, tails are the only outcomes.
• Picture a set of expanding branches. Time 0 is the start. If you flip heads, you move to node H. T otherwise. Flip again, now you have nodes (HH, HT/TH,TT) and so on.
• The values ascribed to H = +1 and T=-1. Profit is simply the sum.
• Do this a gazillion times and at t=10 you will have a normal distribution with mean at zero and standard deviation sqrt (10).
• If you load the coin and give it a 2% greater chance of throwing heads to represent the presence of forecasting power, your distribution at t-10 shifts. The standard deviation hardly moves. That's edge.
• You can insert stops at any of the nodes you want. Say you want to stop on touch of -5. Anything pathway that hits this figure sees you leave the simulation for that round at -5.
• Do that a gazillion times and you will find your distribution skews to the positive, but your expected outcome drops if you have an edge, or does not move if you don't.

That's it. Any single pathway through this garden of forking paths can do whatever it wants. This includes completely defying expectations. If you run the simulations, you will see that there is a reasonable chance of that happening for nearly everything you might reasonably try. What matters for this discussion is whether stops change your return expectations for the better.

As I have said previously:

• They improve your expectations if you have negative edge by taking you out of the game;
• They do nothing if you are random; and
• They reduce expectations if you have edge.

I am not able to replicate your figures. However, the approach you have adopted has strong and likely unintentional bias built in. You seem to be comparing the performance of a portfolio which is an equally weighted one along some measure. You then select a portion in which those which have performed poorly enough along the journey to warrant being stopped out against it. These are removed. What is left is a censored sample which removes all the 'bad' stocks ex-post. Unsurprisingly this generally does a lot better than the uncensored sample. This methodology is unsuitable for the stated purpose of assessing whether stops add value or not on an expectations basis.

 

[tech/a]

Re: Stop Loss is not always your friend....

Excellent posts Craft Know the Future and R/Y.

Some Weekend reading.

 

[craft]

Re: Stop Loss is not always your friend....

Stops do not make money. An edge that creates genuine positive expectancy is the only thing which generates returns on average, through time. Stops do not make money on average through time and reduce your expected profit if you have positive expectancy. In trying to explain this, it is important to separate what makes money and what does not. Edge makes money on average, through time. You will see evidence of belief that stops create expected returns throughout the threads in and of themselves. It cannot.

Risk management is important. I am not doing away with it at all. However, risk management does not make money. It determines magnitude of risk and can alter the shape of your outcomes. You can reshape the distribution via stops (let's not get into mispriced options) to change hit rate, for example. Placing trailing stops will increase hit rate. But it will change the reward profile to offset it in a way that will not change the central expectation. To suggest otherwise is to move into alchemy.

All this pedantry is required here just to divide the role of stops into risk management and/or return generation. It sits squarely in risk management. It has a valuable role to play in regard. I use stops, for example.

RY I understand where you are coming from.

But some positive expectancies are only in existence with a defined price based exit. The stop is required for the positive expectancy to even exist. The stop in this case cannot be considered only in the risk management realm.

To differentiate between what you are saying and where some of the Techs are coming from you need to be able to define the cause of your edge.

It doesn't help in reconciling this discussion that many historical edges identified by only positive expectancy are mined from random distribution skews of no real edge at all and people don't understand that - which is why I am asking what people think the cause of their edge is. Does the cause of your edge require an exit?

Or are we defining stops narrowly here as only arbitrary risk management exits isolated from the edge itself, in which case as you say they can’t be anything but be an expense over the long haul.

 

[DeepState]

Re: Stop Loss is not always your friend....

RY I understand where you are coming from.

But some positive expectancies are only in existence with a defined price based exit. The stop is required for the positive expectancy to even exist. The stop in this case cannot be considered only in the risk management realm.

To differentiate between what you are saying and where some of the Techs are coming from you need to be able to define the cause of your edge.

It doesn't help in reconciling this discussion that many historical edges identified by only positive expectancy are mined from random distribution skews of no real edge at all and people don't understand that - which is why I am asking what people think the cause of their edge is. Does the cause of your edge require an exit?

Or are we defining stops narrowly here as only arbitrary risk management exits isolated from the edge itself, in which case as you say they can’t be anything but be an expense over the long haul.

Yep, we're really on the same page and the rest of this is semantics which you can file as you please.

In the risk/reward dichotomy world, if your stop is directionally informed then I regard it as part of edge. Here's the reason. Let's say you use momentum as your source of edge. Let's say it has positive edge. You can reasonably argue that your momentum signal switches off on exhaustion and know that at a certain price, the signal has no value. You can then place a stop at that point which looks a lot like a trailing stop.

If you are good, is that stop a risk management tool? Or did you know that your insight/edge/yadda shuts down at that point? That was a specific level. It as set with insight. You can regard it as an initiation of a new signal at that point. So it is a stop loss or a take profit or an initiation? Essentially, it is an implementation of an idea with trades accordingly. For example, stronger momentum tends to lead to a stronger signal. Hence you can buy further into the direction. These are stop-initiations. Hardly risk management. Yet they are stops. You are describing the mirror.

If you had no idea and were just using static, none of these ideas would bring expected value to the table.

So, if your stops are directionally informed (as per momentum which is price dependent), I contend that the idea is what creates returns and that's just how you implement them. What is common to all of this is that these stops are of no value if there is no predictive power...which is really all I'm saying here. If you have predictive power over the very long term (like buy-hold style), stops will subtract value. If your insight moves positively with the price (momentum), is it a stop when the signal switches off? Given it is the actual signal that tells you when to switch off, I contend that it isn't the same kind of stop as a stop loss. It's just varying position sizes according to signal strength.

Holy cow, Craft. This is getting really pedantic. I think we both know what we mean in a way which is relevant to us. I think we really are on the same page behind all of this.

 

[craft]

Re: Stop Loss is not always your friend....

I ran a P/B strategy for the last 6 months, with a 5% stop loss and 5% slippage. That beat the same strategy without stop loss by a whopping 16.13%. Attached is a spreadsheet with all the trades, in case anyone is kind enough to look at it and spot an error.

View attachment 60698

Take your largest negative and positive outlier out and then consider both sets of results in light of RY’s post 863.

Also you are introducing a time frame outcome to your system I.e. the exit is no longer just PB>2 or de-listed but the lesser of market price in 6 months or PB>2 for the non stoped system. Introducing this time frame constraint means identifying current momentum is probably going to help and it could be argued the stop does that as the price either has to get on with it or you move on to the next candidate.

The real question is if you let the PB>2 or delisted exit criteria run its full and natural course until everything was exited what would the ultimate return (compound annual return) of that system be compared to running the system with the stop over the same period.

 

[craft]

Re: Stop Loss is not always your friend....

Yep, we're really on the same page and the rest of this is semantics which you can file as you please.

Yep on the same page.

I think if we distinguish between directionally motivated exits from arbitrary exits everybody is on the same page.


That was the point of my posts. Oh and to find out what people think is the 'cause' of the edge they utilise. Any takers?

 

[skyQuake]

Re: Stop Loss is not always your friend....

Then it dawned on me - I was assuming that a 5% stop loss would execute at a price exactly 5% (rounded to nearest $0.01) below purchase price, which is not going to be the case. So, I've put in 5% slippage, each 5% stop loss would actually execute at 10%, 10% at 15%, etc. This brought the result down, but still it is substantially above the benchmark:

View attachment 60697

I tried a High PE strategy, that underperforms the benchmark. I'll spare everyone looking at another graph - the result, again, was that stop losses improved the performance of even an underperforming strategy.

This analysis stands at complete odds with RY's statement that stop losses do not improve the expected return and are simply a risk management mechanism, which is likely to bring down returns, not improve them. When my analysis disagrees with RY, my default stance is that I made a mistake. What did I miss?

RY, how was your data backtested? I ask, because I know standard backtests that re-balance yearly may not be very suitable for testing stop losses.

I ran a P/B strategy for the last 6 months, with a 5% stop loss and 5% slippage. That beat the same strategy without stop loss by a whopping 16.13%. Attached is a spreadsheet with all the trades, in case anyone is kind enough to look at it and spot an error.

View attachment 60698

I think you expanded your stock universe too much. Its including some micro caps that
a) do no volume
b) have incorrect book value

See attached. I've computed avg daily vol in the past 6m, and then taken each position to be the max of (arbitary) $20k or 30% avg daily val, then adjusted each return accordingly.

The avg return is actually very close to 0 (still outperformed the index though)

Eg RDG is your outlier winner. However have a look at the market depth today! Not something you can really take a swing at.

View attachment stoplossTrades_adj.xlsx

 

[tech/a]

Re: Stop Loss is not always your friend....

Once stopped, share will not be bought again the next 365 days.

Why this condition?

Your edge has nothing that indicates momentum in the direction of a positive trade.
Only a statistic.

Net Assets. We buy when P/B < 0.7, we sell when P/B > 2. Is this a proven or hypothetical edge?

Says nothing about its strength in the market NOW.
Does the method buy the stock back in 12 mths regardless of whether it is of less or more value to 12 mths ago.

A simply buy when it rises 10% above todays price should help.

Point is there are so many variables its next to meaningless.

Any takers?

Momentum/Sentiment. That at least gets you on and can be valuable when getting you out---a swing the other way----with momentum.

 

[artist]

Re: Stop Loss is not always your friend....

....

The result comes out of options theory. . . . . here's the argument in a highly simplified framework that actually goes to the binomial lattice methods for options pricing.

. . . .

That's it. Any single pathway through this garden of forking paths can do whatever it wants. This includes completely defying expectations. If you run the simulations, you will see that there is a reasonable chance of that happening for nearly everything you might reasonably try. What matters for this discussion is whether stops change your return expectations for the better.
.

In this context of using stop losses to protect against unacceptable, unexpected (large) moves in the SP, isn't the appropriate distribution the Poisson Distribution rather than the Binomial ?


"The Black-Scholes Model
l
The binomial model is a discrete-time model for asset price
movements, with a time interval (t) between price movements.
l
As the time interval is shortened, the limiting distribution, as t -> 0,
can take one of two forms.
–
If as t -> 0, price changes become smaller, the limiting distribution is the
normal distribution and the price process is a continuous one.
–
If as t->0, price changes remain large, the limiting distribution is the
poisson distribution, i.e., a distribution that allows for price jumps.
l
The Black-Scholes model applies when the limiting distribution is the
normal distribution , and explicitly assumes that the price process is
continuous and that there are no jumps in asset prices." (http://people.stern.nyu.edu/adamodar/pdfiles/option.pdf on Page 12)


and

"The Poisson distribution can be applied to systems with a large number of possible events, each of which is rare. How many such events will occur during a fixed time interval? Under the right circumstances, this is a random number with a Poisson distribution.

. . .

Applications of the Poisson distribution can be found in many fields related to counting:

. . .

The number of jumps in a stock price in a given time interval.

. . . " (http://en.wikipedia.org/wiki/Poisson_distribution)

 

[DeepState]

Re: Stop Loss is not always your friend....

In this context of using stop losses to protect against unacceptable, unexpected (large) moves in the SP, isn't the appropriate distribution the Poisson Distribution rather than the Binomial ?


"The Black-Scholes Model
l
The binomial model is a discrete-time model for asset price
movements, with a time interval (t) between price movements.
l
As the time interval is shortened, the limiting distribution, as t -> 0,
can take one of two forms.
–
If as t -> 0, price changes become smaller, the limiting distribution is the
normal distribution and the price process is a continuous one.
–
If as t->0, price changes remain large, the limiting distribution is the
poisson distribution, i.e., a distribution that allows for price jumps.
l
The Black-Scholes model applies when the limiting distribution is the
normal distribution , and explicitly assumes that the price process is
continuous and that there are no jumps in asset prices." (http://people.stern.nyu.edu/adamodar/pdfiles/option.pdf on Page 12)


and

"The Poisson distribution can be applied to systems with a large number of possible events, each of which is rare. How many such events will occur during a fixed time interval? Under the right circumstances, this is a random number with a Poisson distribution.

. . .

Applications of the Poisson distribution can be found in many fields related to counting:

. . .

The number of jumps in a stock price in a given time interval.

. . . " (http://en.wikipedia.org/wiki/Poisson_distribution)

Black Scholes assumes log-normal distribution of returns. This lattice is modeling along those lines. The lattice approach was invented by Merton and the approach produces returns that asymptote to the Black Scholes formulation. Binomial converges to normal at the asymptote. It's bloody amazing that central limit stuff. Poisson can be suitable where you do not know the variance. In this case you do.

 

[artist]

Re: Stop Loss is not always your friend....

Black Scholes assumes log-normal distribution of returns. This lattice is modeling along those lines. The lattice approach was invented by Merton and the approach produces returns that asymptote to the Black Scholes formulation. Binomial converges to normal at the asymptote. It's bloody amazing that central limit stuff. Poisson can be suitable where you do not know the variance. In this case you do.

I don't want to delve too far into option pricing models in this thread. But obviously volatility is the issue that gives rise to market participants trying to devise stop-loss strategies, yet not get stopped out so often that the practice becomes a source of loss instead in its own right.

With regard to B-S / Binomial / Normal / Lognormal, one of my finance textbooks has this to say; "[A]s the number of subintervals (or nodes in the lattice in your terminology) increases, the number of possible stock prices also increase. . . . . [and] the graph approaches the appearance of the familiar bell-shaped curve. In fact, as the number of intervals increases . . . the frequency distribution progressively approaches the lognormal distribution rather than the normal distribution."

That is all well and good, but there is a footnote attached to the last sentence. "Actually, more complex considerations enter here. The limit of this process is lognormal only if we assume also that stock prices move continuously, by which we mean that over small time intervals only small price movements can occur. This rules out rare events such as sudden,extreme price moves in response to dramatic information (like a takeover attempt)."

Moreover, the same textbook lists three "important assumptions underlying the [Black-Scholes] formula". The third of the three they list is "Stock prices are continuous, meaning that sudden extreme jumps such as those in the aftermath of an announcement of a takeover attempt are ruled out."

That is, the models rule out of consideration precisely the fluctuations that this thread is discussing.

You state that "In this case you do (know the variance)". But you only know the historical variance over an arbitrarily selected period of time. Choose a different data interval and you will get a different historical variance figure. There are different ways to try to model forward volatility estimates (e.g. (G)ARCH http://en.wikipedia.org/wiki/Autoregressive_conditional_heteroskedasticity) but no method of assessing historical, or estimating future volatility can give you confidence that a sudden short-term change in price (spike or otherwise) will not adversely affect your position.

So a long-term investor may well be able to ride out a short-term volatility spike. So might a day-trader. It will depend on the usual things such as risk-tolerance, money management etc. The usual. Including, perhaps, the use of a stop-loss.

 

[DeepState]

Re: Stop Loss is not always your friend....

I don't want to delve too far into option pricing models in this thread. But obviously volatility is the issue that gives rise to market participants trying to devise stop-loss strategies, yet not get stopped out so often that the practice becomes a source of loss instead in its own right.

With regard to B-S / Binomial / Normal / Lognormal, one of my finance textbooks has this to say; "[A]s the number of subintervals (or nodes in the lattice in your terminology) increases, the number of possible stock prices also increase. . . . . [and] the graph approaches the appearance of the familiar bell-shaped curve. In fact, as the number of intervals increases . . . the frequency distribution progressively approaches the lognormal distribution rather than the normal distribution."

That is all well and good, but there is a footnote attached to the last sentence. "Actually, more complex considerations enter here. The limit of this process is lognormal only if we assume also that stock prices move continuously, by which we mean that over small time intervals only small price movements can occur. This rules out rare events such as sudden,extreme price moves in response to dramatic information (like a takeover attempt)."

Moreover, the same textbook lists three "important assumptions underlying the [Black-Scholes] formula". The third of the three they list is "Stock prices are continuous, meaning that sudden extreme jumps such as those in the aftermath of an announcement of a takeover attempt are ruled out."

That is, the models rule out of consideration precisely the fluctuations that this thread is discussing.

You state that "In this case you do (know the variance)". But you only know the historical variance over an arbitrarily selected period of time. Choose a different data interval and you will get a different historical variance figure. There are different ways to try to model forward volatility estimates (e.g. (G)ARCH http://en.wikipedia.org/wiki/Autoregressive_conditional_heteroskedasticity) but no method of assessing historical, or estimating future volatility can give you confidence that a sudden short-term change in price (spike or otherwise) will not adversely affect your position.

So a long-term investor may well be able to ride out a short-term volatility spike. So might a day-trader. It will depend on the usual things such as risk-tolerance, money management etc. The usual. Including, perhaps, the use of a stop-loss.

The above relates to the discussion on options pricing assumptions. The assumption of log normality and continuous markets is violated in reality and the actual pricing of options reflects this (via skews in the implied volatilities through time and across time). Nonetheless, without any directional prediction ability, the placement of stops even in that environment or any other arbitrary distribution whose expected return is zero and will not see expectations shift through the use of stops alone. If it did, it would be the result of a coding error.

The lattice allows for the fact that stocks trade in a non-continuous way. Stocks trade by the tick, for example. You can also vary heaps of things including distribution standard deviation, skew, kurtosis and higher moments, E/F/I/etc.-GARCH / Regime shift, jump models, pairs can include copula, resample in any way, distort these via generation functions etc..probabilities through each node path etc.. Anything which moves can be modeled including take-overs and extinctions. Anything that you have raised is easily captured. And then some. The coin flip at +1/-1 is clearly the most basic of basic, but it is sufficient for the purpose. Adding infinite model features doesn't change the outcome. This is just a model to do a brute force simulation on what is a tautology that actually doesn't need simulation to prove it exists.

Alpha is zero sum prior to costs. Given it is zero sum, someone with no insight can't suddenly generate returns by whacking down stops. A second person will eventually meet them on a forum and then whack down stops too. They'll all do it and then suddenly everyone with no idea what is going on is making money in a sub-zero sum game. Suddenly the most reliable thing to do is to have no idea but whack down stops. More stops the better. That's alchemy. Stops do not create returns in and of themselves on average through time.

We are talking about expected returns and stop losses. We are using a framework lifted from the options world because it is suitable. Adding stops will simply alter the distribution outcomes. It may make you feel safer, keep you in the game longer perhaps, but not ultimately increase your expected return if you have no edge. I have already covered off what happens if you do have a positive or negative edge. These continue to apply in this generalized environment which captures everything that a market can do.

Feel free to model something and confirm the result. Any distributions assumptions are fine but the expected return must be zero. Add stops as you please as long as they are not directionally informed. That's the only two conditions. Zero mean distribution, stops not directionally informed. ...then see what happens. It's already pre-determined. There will be no move in expected outcome although the distribution will shift in some way depending on your other assumptions - which, whilst interesting and valid, don't matter for the outcome on the question of whether stops add value in and of themselves on average through time.

Once again, that is not to say stops do not have value as risk management tools. But, that's risk management.

 

[tech/a]

Re: Stop Loss is not always your friend....

Too much pontification

 

[DeepState]

Re: Stop Loss is not always your friend....

Too much pontification

Tends to happen when tautologies are not recognised as such and when things that aren't are thought to be so. A tautology requires no pontification. It just is.

I guess we agree. Apparently a stop loss is not always your friend.

 

[KnowThePast]

Re: Stop Loss is not always your friend....

....

The result comes out of options theory. It's fairly straight forward. Given your quant skills and for those inclined to stochastic thinking, here's the argument in a highly simplified framework that actually goes to the binomial lattice methods for options pricing.

• You have a fair coin (ie. no forecasting power).
• We flip this coin 10 times. Heads, tails are the only outcomes.
• Picture a set of expanding branches. Time 0 is the start. If you flip heads, you move to node H. T otherwise. Flip again, now you have nodes (HH, HT/TH,TT) and so on.
• The values ascribed to H = +1 and T=-1. Profit is simply the sum.
• Do this a gazillion times and at t=10 you will have a normal distribution with mean at zero and standard deviation sqrt (10).
• If you load the coin and give it a 2% greater chance of throwing heads to represent the presence of forecasting power, your distribution at t-10 shifts. The standard deviation hardly moves. That's edge.
• You can insert stops at any of the nodes you want. Say you want to stop on touch of -5. Anything pathway that hits this figure sees you leave the simulation for that round at -5.
• Do that a gazillion times and you will find your distribution skews to the positive, but your expected outcome drops if you have an edge, or does not move if you don't.

View attachment 60699

That's it. Any single pathway through this garden of forking paths can do whatever it wants. This includes completely defying expectations. If you run the simulations, you will see that there is a reasonable chance of that happening for nearly everything you might reasonably try. What matters for this discussion is whether stops change your return expectations for the better.

Doesn't that assume each step to have an equal probability of happening? Whilst these are all possilities, will real world behave like that?

I am not able to replicate your figures.

Just to confirm, you ran the same backtest? (if so, a code bug in my procedure is more likely.)

My universe is not entire XAO, so that could be a reason for the difference as well.

However, the approach you have adopted has strong and likely unintentional bias built in. You seem to be comparing the performance of a portfolio which is an equally weighted one along some measure. You then select a portion in which those which have performed poorly enough along the journey to warrant being stopped out against it. These are removed. What is left is a censored sample which removes all the 'bad' stocks ex-post. Unsurprisingly this generally does a lot better than the uncensored sample. This methodology is unsuitable for the stated purpose of assessing whether stops add value or not on an expectations basis.

I am not sure I follow. The universe was the same for all tests, stop loss runs and benchmarks runs were identical, other than the stop loss. Yes, it is not a suitable methodology to test every possibility, but that wasn't the aim - I just wanted to see what has actually happened over the last 5 years.

Take your largest negative and positive outlier out and then consider both sets of results in light of RY’s post 863.

Also you are introducing a time frame outcome to your system I.e. the exit is no longer just PB>2 or de-listed but the lesser of market price in 6 months or PB>2 for the non stoped system. Introducing this time frame constraint means identifying current momentum is probably going to help and it could be argued the stop does that as the price either has to get on with it or you move on to the next candidate.

The real question is if you let the PB>2 or delisted exit criteria run its full and natural course until everything was exited what would the ultimate return (compound annual return) of that system be compared to running the system with the stop over the same period.

Hi craft,

But that's exactly what I did, I think

My "no stop loss" run is the PB>2 exit critera allowed to run its course.
I then re-ran it with various stop losses, and these all showed an improved result over the "no stop loss" run.

Regarding the outliers, these are also present in "no stop loss" benchmark, so they do not explain the outperformance.

That was the point of my posts. Oh and to find out what people think is the 'cause' of the edge they utilise. Any takers?

I'll bite - size, liquidity and psychology. Not necessarily in that order.

I think you expanded your stock universe too much. Its including some micro caps that
a) do no volume
b) have incorrect book value

See attached. I've computed avg daily vol in the past 6m, and then taken each position to be the max of (arbitary) $20k or 30% avg daily val, then adjusted each return accordingly.

The avg return is actually very close to 0 (still outperformed the index though)

Eg RDG is your outlier winner. However have a look at the market depth today! Not something you can really take a swing at.

View attachment 60703

Thanks skyQuake!

These are all valid points. But, all these are present in the "no stops" benchmark as well. So outperformance is not explained by these.

I re-ran the tests, this time only including stocks with market cap over $500m, which should all have enough liquidity. Same result.

Why this condition?

Your edge has nothing that indicates momentum in the direction of a positive trade.
Only a statistic.

Net Assets. We buy when P/B < 0.7, we sell when P/B > 2. Is this a proven or hypothetical edge?

Proven. Low P/B strategies have been shown to consistently outperform the market.

The condition of 365 days is to stop the stock from been bought immediately after getting stopped out (as it would still meet the buy criteria). I didn't want to put more criteria in, as I was trying to just test the effect of stop losses. Playing around with duration or other criteria, including momentum, could certainly improve the results.

Says nothing about its strength in the market NOW.
Does the method buy the stock back in 12 mths regardless of whether it is of less or more value to 12 mths ago.

A simply buy when it rises 10% above todays price should help.

Point is there are so many variables its next to meaningless.

Buys back regardless. I didn't mean it as a good strategy, just a suitable example to check how stop losses would have effected it over the last 5 years.

Alpha is zero sum prior to costs. Given it is zero sum, someone with no insight can't suddenly generate returns by whacking down stops. A second person will eventually meet them on a forum and then whack down stops too. They'll all do it and then suddenly everyone with no idea what is going on is making money in a sub-zero sum game. Suddenly the most reliable thing to do is to have no idea but whack down stops. More stops the better. That's alchemy. Stops do not create returns in and of themselves on average through time.

First of all - I fully agree with you. This is alchemy, so please don't take it that I am disagreeing with you, I am trying to make it go away in my data

Why is it that Low PE/PB strategies have outperformed for such a long time? Yes, their effectiveness has diminished, but not zeroed out as one would expect, as everyone joins the party? Size and liquidity issues more likely for these stocks, perhaps?

I now ran more backtests on more strategies. I also googled many studies on this. The results are mixed, but certain trends can be "mined".

- Performance tends to be improved when using static (EMH), or simple mean reversion strategies (fundamental).
- The strategies that significantly outperform the benchmark are not helped by stop losses.
- The larger the outperformance, the more stop losses take away from it.
- strategies with large underperformance have been helped by stop losses.
- strategies with relatively small over/under performance, even if consistent, have mixed results.

So, now that I've done more runs, the results seem to be exactly what you were suggesting:
- stop losses worsen returns for highly profitable strategies (edge?)
- stop losses improve returns for bad strategies
- results in the middle are highly mixed.

All that work to find out that I should have just listened to you. You've mentioned before that 95% of the signals you find turn out to be just noise - this has been exactly my experience. Add one to the list.

Getting on to risk - do stop losses reduce risk in a highly diversified portfolio? I would expect that they change the shape of losses, but not the final outcome.

 

[luutzu]

Re: Stop Loss is not always your friend....

All these reminds me of a story I heard somewhere. It goes something like this:

Two country Gentlemen at a bar were having philosophical debates and somehow it got to them debating how many teeth does a horse in the barn outside have.

One said that horses are of that genus, related to this and that; and this and that have this many teeth... horses being bigger and eat grass and at certain age it have this many teeth;

The other argued that it depends on the origin of the horse in question... Arabian horses would have this many teeth, factor in the climate and this and that, it would have this many teeth at this and that age, depends on age and birth and health blah blah.

They debated back and forth, back and forth into the early hours... still keep going until a lowly, uneducated bartender told them...

The barn's unlocked, why don't you guys just go out there and open the horse's mouth and start counting.



---------

With investing, you don't have a lowly, uneducated bartender telling you... you got at least two self-made multi-billionaire investors telling you to go and open the horse's mouth and start counting. But somehow real, smart, investing just doesn't work like that.

 

[KnowThePast]

Re: Stop Loss is not always your friend....

All these reminds me of a story I heard somewhere. It goes something like this:

Two country Gentlemen at a bar were having philosophical debates and somehow it got to them debating how many teeth does a horse in the barn outside have.

One said that horses are of that genus, related to this and that; and this and that have this many teeth... horses being bigger and eat grass and at certain age it have this many teeth;

The other argued that it depends on the origin of the horse in question... Arabian horses would have this many teeth, factor in the climate and this and that, it would have this many teeth at this and that age, depends on age and birth and health blah blah.

They debated back and forth, back and forth into the early hours... still keep going until a lowly, uneducated bartender told them...

The barn's unlocked, why don't you guys just go out there and open the horse's mouth and start counting.



---------

With investing, you don't have a lowly, uneducated bartender telling you... you got at least two self-made multi-billionaire investors telling you to go and open the horse's mouth and start counting. But somehow real, smart, investing just doesn't work like that.

What you are missing is how much fun these two guys had. And that one horse could have been an outlier, the results from it don't count

On a serious note, luutzu, a lot of you posts come across as advocating ignorance as a strategy. While many things are unnecessarily complicated, it doesn't make all that is complicated wrong.