The Portfolio MYTH-----Do you need to change your thinking?

[Quant]

You need to watch for the offshore market activity now who don't get credits. They are increasingly crushing the trade before ex. That is their way of playing this arb.

Yes correct , something im aware of and has been the case as long as i can remember , i never hold into EX div or report for these very reasons , in fact under the right circumstance ive been known to short stocks pre report , obv certain parameters required , i dont short stocks often though

 

[Quant]

Can you please expand on this? This looks important to your perspective.

in a 30 stock portfolio with lets say equal weight a stock going to zero is a 3% hit to portfolio , in a 10 stock port its 10% obviously , you cant bury the risk , you are made to deal with it and thats a positive thing imo . Now we get index correlation on 30 stocks that we dont deal with the risk and by the time we panic the drawdown is 30% ( hypothetical numbers ) diversification is useless in deep index drawdowns

 

[DeepState]

in a 30 stock portfolio with lets say equal weight a stock going to zero is a 3% hit to portfolio , in a 10 stock port its 10% obviously , you cant bury the risk , you are made to deal with it and thats a positive thing imo . Now we get index correlation on 30 stocks that we dont deal with the risk and by the time we panic the drawdown is 30% ( hypothetical numbers ) diversification is useless in deep index drawdowns

Totally fair point. If we rely too much on statistical arguments on diversification, then we will be inevitably disappointed when the diversification benefits vapourise just when we need them. Although we can allow for these effects by using 'stressed' matrices in port con for example.

Can I take you up on that 'can't bury the risk'. Just because I have more at risk doesn't necessarily make me a whole lot better at managing it. Though some have argued that focus increases ability to generate return and, probably, understand risk better, the extent to which this occurs in practice is not normally very large. So, beyond putting stops etc. are you inferring that we somehow are able to be more aware of risks and doing smart things which we would not otherwise do? And, net, this gets a better outcome? Using a ten stock portfolio vs, say, a 30 stock one as comparisons for concrete examples (feel free to use others though).

 

[Quant]

Can I take you up on that 'can't bury the risk'. Just because I have more at risk doesn't necessarily make me a whole lot better at managing it. Though some have argued that focus increases ability to generate return and, probably, understand risk better

Well a skill set is obviously required to deal with the risk and goes without saying a skill cant be gained/improved if you dont use/ strategize it . My theory is the smaller portfolio in a way forces you to go down this road . Also having a small universe allows you to be more active and knowledgable on the stocks you buy . This is why i prefer large caps , not so much smoke and mirrors and the times to deal with unexpected results is very very low , this is part of why i wont hold into reports , this is when risk is likely announced . Give me the easy money anyday over the maybe money

 

[DeepState]

Well a skill set is obviously required to deal with the risk and goes without saying a skill cant be gained/improved if you dont use/ strategize it . My theory is the smaller portfolio in a way forces you to go down this road . Also having a small universe allows you to be more active and knowledgable on the stocks you buy . This is why i prefer large caps , not so much smoke and mirrors and the times to deal with unexpected results is very very low , this is part of why i wont hold into reports , this is when risk is likely announced . Give me the easy money anyday over the maybe money

OK. So it is mostly true that, if you have a stack of stocks, almost any reasonable weighting produces a half-decent risk outcome. You don't have to think. However, if you are aware of risk, you have to go at it pretty hard when individual positions matter. Very sound. I agree. Beyond about 15-20 positions and, you've done the thinking, portfolio risk gains from further diversification is pretty limited in absolute terms...but still very present if index relative.

Can I take it that you aren't in favour of a two-stock portfolio but find something like 10-20 appropriate (if fully deployed)? You are clearly risk aware. That's still punchy but if drawdown won't smash your lifestyle, and your horizon is long, and you don't answer to anyone else...it's about right for most people in that situation..I think. Unless the whole lot is in specs etc etc.

 

[howardbandy]

Thanks for the clarification.

So, for a situation there there is even a low (but strictly positive) degree of aversion to loss and even high predictive ability (say, Info Coefficient of 0.2 at monthly interval), how many stocks might appear when the universe is the ASX 200 or S&P500? And signals are available on the full cross-section.....

Would it be 2? Or closer to 20+

Greetings --

The idea is to develop systems for one issue at a time. Each system trades one issue long/flat or, if you wish, short flat -- but short systems are harder to develop and many people do not / cannot go short.

At each evaluation and management period -- which I recommend to be daily -- update all the systems and compute the CAR25 for each of them based on the "shadow" trades that they would have taken. Following safe-f, use the available funds to take one position based on the one system that has the highest CAR25. If the highest CAR25 of all the systems tracked is $20 bills under the mattress, remain flat for that day.

Best, Howard

 

[ducati916]

Morning chaps,
A couple of links for you:

http://blog.alphaarchitect.com/2017...cker-of-all-time-buffett-or-lynch/#gs.bg6Rej0

https://www.ifa.com/articles/what_about_active_managers_have_beaten_market_still_need_alpha_charts/

jog on
duc

 

[ducati916]

Chaps,

With regard to the original question:

If you are day-trading, entry/exit within 1 trading session, you can still be 'diversified' even if you only ever trade the ES contract (example).

So you in this scenario would have capital of $50K. You have a couple of options:

(a) you take a single trade for the day, using 100% of your $50K (and any margin that you may qualify for), and exit for a profit/loss before the close; or

(b) you have the same $50K, but, you will enter any number of trades in a given trading session, closing flat at the close. In this scenario you only allocate a % of capital to any given trade.

Option (b) is a diversified 'portfolio' in that you have a trigger (expectancy) for opening a trade and closing that trade if another signal (stop) is triggered. Your expectancy is that there will be winners and losers, but overall, you will over time (day, week, month) be profitable. You are diversified in that you will have X number of trades on any given day.

Option (a)'s expectancy should be much higher than Option (b) as you are only taking 1 trade (no diversification). If you have a high success rate, this is obviously a very profitable strategy.

Anything other than a 'day-trade' is (in my book) an investment, as, when the market closes and you can no longer place a trade to exit/hedge your position, you are vulnerable to 'news', whether that is for 2 days or 20 years.

To my mind, the lower your diversification in this scenario, the more risk you are assuming.

With ETF's, you can have a single position to manage, but with an already built in diversification. You will give up the rocket ship, but you won't be caught by the liquidators announcing entry to your stock either.

jog on
duc

 

[DeepState]

Greetings --

The idea is to develop systems for one issue at a time. Each system trades one issue long/flat or, if you wish, short flat -- but short systems are harder to develop and many people do not / cannot go short.

At each evaluation and management period -- which I recommend to be daily -- update all the systems and compute the CAR25 for each of them based on the "shadow" trades that they would have taken. Following safe-f, use the available funds to take one position based on the one system that has the highest CAR25. If the highest CAR25 of all the systems tracked is $20 bills under the mattress, remain flat for that day.

Best, Howard

So a single issue at any time with allocation at 'safe-f'? One stock portfolio (or cash)?

The concepts of diversification at the portfolio level come from time diversification and also from stochastic, Bayesian, position sizing to control risk exposure.

Greetings --
Post # 1 of this thread. Tech/a is correct. End of thread!!

So this was the literal perspective. Yes, it seemed strange for you to write something you didn't actually mean but it seemed so out of step with conventional belief I wasn't sure quite what to make of it.

----

This is very divergent which is absolutely fantastic to find.

Let's say TraderX has a preferred habitat consisting of the universe of stocks in the ASX 200 + Russell 2000. TraderX is able to generate 1000+ candidate longitudinal signals in each of these 2,200 stocks. These need not be using the same functional form (albeit with different parameterisation). You can think of it as 2,200x1000+ candidate signals. This is, after all, machine learning within an information rich dataset at reasonably high frequency.

Let's make a huge assumption that TraderX has investment (ok, trading, for this conversation) insight. In other words, those signals actually have a basis for prediction and are not formed via extensive mining of random or uninformative data. Each is genuinely believed to have true predictive ability, albeit time-varying.

Out of 2.2million+ signals, we only choose the best one to invest in for that day?? Would the second best signal (which would have been the best had not some freak of nature actual found the better one) not be valid at all? Signal #1 has valid forecasting, but signal#2 to 2,200,000+ have no validity? Yes, some will be short selling etc, but let's not move to the edge cases yet. Surely one of 2.2million signals will pick a buy for one stock even if that signal might be #10 on the day...Signal ranked 10 of 2.2million is given zero weight?

If signal #1 has scaled 'merit' of 100%, even if signal #2 (on CAR25) has 80% scaled merit and signal #3 has 60% etc... All key risk features and curve characteristics would improve for compiling these in to a portfolio and weighting in a way that makes sense...say using safe-f for a portfolio of these particular signals in some sensible way.

If TraderX has produced 2.2million candidate signals and is skilled, it would be the most unbelievable situation for them to believe there was literally only one idea that was worth investing in that day and lob their entire risk tolerance on to it.


So..unless the humble CAR25 is some magic filter which finds the one valid needle that day out of a veritable haystack, whose composition changes from hay to needle and back in a non-linear way by the day, I must be misunderstanding another heterodox concept. Can you please clarify what that is on this occasion?

The process is undeniably optimal if the insights are random but risk can be estimated in that it cuts transactions costs vs other portfolio oriented alternatives for a given risk tolerance. However, I think your ambitions for it are much higher than that ('towards best').

 

[InsvestoBoy]

So...I must be misunderstanding another concept. Can you please clarify what that is on this occasion?

From my understanding of what Dr Bandy is saying:
You have a stable of systems that exploit different patterns and market regimes.

Over the intermediate term one market pattern or regime becomes dominant, and the system that exploits that will float to the top. I doubt the systems are allocated to a single stock, probably a ranking of stocks that match the regime best or a broad ETF or futures contract (which is already a portfolio, NO FOOLING).

So on any given day you are likely "investing" into the same signal that you did yesterday or the day before that (and probably the same signal tomorrow).

e.g. if buying ES on 2 day closing low and selling on the first up close is profitable, then you will be doing that until another regime becomes dominant.

 

[DeepState]

I doubt the systems are allocated to a single stock, probably a ranking of stocks that match the regime best or a broad ETF or futures contract (which is already a portfolio, NO FOOLING).

So on any given day you are likely "investing" into the same signal that you did yesterday or the day before that (and probably the same signal tomorrow).

e.g. if buying ES on 2 day closing low and selling on the first up close is profitable, then you will be doing that until another regime becomes dominant.

That's more reasonable in my view, yet Howard's own words (which are usually very accurately chosen) say: "Following safe-f, use the available funds to take one position based on the one system that has the highest CAR25"

One system. One stock. Yes, that stock may be a diversified ETF. I imagine that the process is also intended to be applied to individual stocks and is more generally expressed or applied for such processes or trading across a universe of indices. So, one instrument selected for one system for full risk exposure, per day.

 

[howardbandy]

That's more reasonable in my view, yet Howard's own words (which are usually very accurately chosen) say: "Following safe-f, use the available funds to take one position based on the one system that has the highest CAR25"

One system. One stock. Yes, that stock may be a diversified ETF. I imagine that the process is also intended to be applied to individual stocks and is more generally expressed or applied for such processes or trading across a universe of indices. So, one instrument selected for one system for full risk exposure, per day.

Greetings --

Deepstate is stating my recommendation correctly. To reiterate the process:

1. Use the analysis of the price history of a list of potentially tradable issues to determine the risk, profit potential, and liquidity of each. Issues that pass those filters become candidates to be traded. They might be individual stocks, individual futures contracts, ETFs that are made up of some group of issues. This is the activity of the "data prospector." No rules (no model) have yet been applied. Independent of any possible rules, some data series are untradable due to high or irregular volatility, others are not worth trading due to low volatility. The individual issues can be given a score to rank them.

2. One at a time, pick a trading candidate and develop a model (a set of rules that identifies probably profitable trades and issues orders to buy and to sell) that trades that single issue long and flat. Or short and flat, but not both in a single model. For stocks and equity ETFs, identifying good signals to be short is more difficult than to be long. Patterns preceding price rises are different than patterns preceding price falls. Do not ask your model to try to identify both and keep them separate.

The process of system development becomes the process of fitting a model to a single data series. It is during this process that we discover whether there are identifiable and persistent signals in the data that precede profitable trades. Not every candidate data series that the prospector gave a passing grade to results in a tradable system.

Use the scientific method. Examine historical data looking for patterns and fit the model to the data, giving the best in-sample fit. Test the model on data not previously used in fitting (validation). The test produces a set of trades that are the best estimate of future performance. Ignore in-sample trade results -- they have no value in estimating future performance. Using the best estimate set of trades, compute metrics giving the maximum safe position size (safe-f) and estimate of compound annual rate of return (CAR25). If this one-time validation shows satisfactory results -- that is a value of CAR25 that is higher than risk free use of the funds -- move the system to trading where it will be managed by a trading management model.

Develop as many such systems as you wish. Each has its own safe-f and CAR25.

3. The trading management model uses trades as its input and has safe-f and CAR25 as its output. The model is based on Bayesian analysis and sequential learning -- learn from the most recent data, adjusting your view as new data is received. Begin with the best estimate set of trades for any system under consideration. Trade, or paper trade, that system producing real trades if real money was used or shadow trades if paper traded. Add the new trades to the best estimate set, reducing the weighting of older trades as new trades are added. Safe-f and CAR25 will change. The change represents changes in the data series. We cannot know whether these are random and within the distribution we began with, or "regime change" coming from a new and different distribution. We act as if they are true indications of distribution shift in the signals coming from the data. Recompute safe-f and CAR25 at every evaluation -- probably at the close of trading every day -- and sort the results. The system to trade next is the one at the top of the list. That is, trade the one system that has the highest risk-normalized CAR. Take a position guided by safe-f. To trade at any position size higher than safe-f risks drawdown that exceeds your tolerance.

Diversity comes from having multiple systems, each of which trades a single issue, available for trading at the next opportunity. As conditions change, risk changes, safe-f changes, CAR25 changes, leadership of the list changes, and the system that is being traded changes -- or, at least, could change.

The portfolio is a portfolio of systems, each trading a single issue with its own set of rules -- rather than a portfolio of issues traded by a commonly applied single set of rules. In terms of modern portfolio theory and the efficient frontier, the chart is one dimensional -- CAR25 along a line -- rather than two dimensional -- risk (typically standard deviation) on one axis, return (typically CAR) on the other. The method I recommend has normalized the risk, so risk is the same for all system. The other dimension is return and that is CAR25.

Best regards, Howard

 

[craft]

Greetings --

Deepstate is stating my recommendation correctly. To reiterate the process:

1. Use the analysis of the price history of a list of potentially tradable issues to determine the risk, profit potential, and liquidity of each. Issues that pass those filters become candidates to be traded. They might be individual stocks, individual futures contracts, ETFs that are made up of some group of issues. This is the activity of the "data prospector." No rules (no model) have yet been applied. Independent of any possible rules, some data series are untradable due to high or irregular volatility, others are not worth trading due to low volatility. The individual issues can be given a score to rank them.

2. One at a time, pick a trading candidate and develop a model (a set of rules that identifies probably profitable trades and issues orders to buy and to sell) that trades that single issue long and flat. Or short and flat, but not both in a single model. For stocks and equity ETFs, identifying good signals to be short is more difficult than to be long. Patterns preceding price rises are different than patterns preceding price falls. Do not ask your model to try to identify both and keep them separate.

The process of system development becomes the process of fitting a model to a single data series. It is during this process that we discover whether there are identifiable and persistent signals in the data that precede profitable trades. Not every candidate data series that the prospector gave a passing grade to results in a tradable system.

Use the scientific method. Examine historical data looking for patterns and fit the model to the data, giving the best in-sample fit. Test the model on data not previously used in fitting (validation). The test produces a set of trades that are the best estimate of future performance. Ignore in-sample trade results -- they have no value in estimating future performance. Using the best estimate set of trades, compute metrics giving the maximum safe position size (safe-f) and estimate of compound annual rate of return (CAR25). If this one-time validation shows satisfactory results -- that is a value of CAR25 that is higher than risk free use of the funds -- move the system to trading where it will be managed by a trading management model.

Develop as many such systems as you wish. Each has its own safe-f and CAR25.

3. The trading management model uses trades as its input and has safe-f and CAR25 as its output. The model is based on Bayesian analysis and sequential learning -- learn from the most recent data, adjusting your view as new data is received. Begin with the best estimate set of trades for any system under consideration. Trade, or paper trade, that system producing real trades if real money was used or shadow trades if paper traded. Add the new trades to the best estimate set, reducing the weighting of older trades as new trades are added. Safe-f and CAR25 will change. The change represents changes in the data series. We cannot know whether these are random and within the distribution we began with, or "regime change" coming from a new and different distribution. We act as if they are true indications of distribution shift in the signals coming from the data. Recompute safe-f and CAR25 at every evaluation -- probably at the close of trading every day -- and sort the results. The system to trade next is the one at the top of the list. That is, trade the one system that has the highest risk-normalized CAR. Take a position guided by safe-f. To trade at any position size higher than safe-f risks drawdown that exceeds your tolerance.

Diversity comes from having multiple systems, each of which trades a single issue, available for trading at the next opportunity. As conditions change, risk changes, safe-f changes, CAR25 changes, leadership of the list changes, and the system that is being traded changes -- or, at least, could change.

The portfolio is a portfolio of systems, each trading a single issue with its own set of rules -- rather than a portfolio of issues traded by a commonly applied single set of rules. In terms of modern portfolio theory and the efficient frontier, the chart is one dimensional -- CAR25 along a line -- rather than two dimensional -- risk (typically standard deviation) on one axis, return (typically CAR) on the other. The method I recommend has normalized the risk, so risk is the same for all system. The other dimension is return and that is CAR25.

Best regards, Howard

In summary, it appears you are describing continuously changing to the system with the best historical risk adjusted return.


How different is that from continuously changing to the fund manager with the best historical risk adjusted return? The inclination to chase return seems ingrained to human nature so plenty of historical data on the success of such strategy exists. The evidence of its success is overwhelmingly negative.


Obviously you are talking changing systems not fund managers, although given fund managers are normally pretty sticky to one system /strategy this distinction may not be so great. And the rate of changing systems you advocate may be greater than what has been studied in the Phenomena of chasing historical fund manager returns. I would like to see any ex-post evidence you can provide that your twists on this human trait to chase historical returns will produce results opposite to the body of evidence that already exists?

 

[howardbandy]

In summary, it appears you are describing continuously changing to the system with the best historical risk adjusted return.


How different is that from continuously changing to the fund manager with the best historical risk adjusted return? The inclination to chase return seems ingrained to human nature so plenty of historical data on the success of such strategy exists. The evidence of its success is overwhelmingly negative.


Obviously you are talking changing systems not fund managers, although given fund managers are normally pretty sticky to one system /strategy this distinction may not be so great. And the rate of changing systems you advocate may be greater than what has been studied in the Phenomena of chasing historical fund manager returns. I would like to see any ex-post evidence you can provide that your twists on this human trait to chase historical returns will produce results opposite to the body of evidence that already exists?

Hi Craft --

Yes. Use whatever model currently gives the best prediction of future performance, as measured by metrics appropriate for the application. This is the central principle of sequential learning and adaptive management. Credit scoring, self-driving cars, paper machine operation, etc. And trading.

There are variations of changing mutual funds and exchange traded funds that do work when rapid switching and short holding periods are allowed, easy, and inexpensive.

The literature describes rotational systems, often based on mutual funds. Working with different fund managers complicates use of traditional mutual funds. Mutual funds try to avoid customers rapid switching in and out, often imposing early-termination fees to discourage it and overwhelm whatever profits are made.

It isn't risk adjusted long term historical return that is most useful to switching -- it is recent trade performance. Day-by-day. Simplistic models may not work at all. There must be reasonable assurance, through use of sound modeling and simulation techniques, that the model applied to the data series does identify profitable patterns. Not all data series qualify. Not all models work.

The literature I have seen that describes historical fund manager performance and the value of chasing it uses long lookback periods, long holding periods, and no scientific method. The conclusions are, as you say, disappointing.

A technique that does work is to use the mutual funds to generate signals, then take the trades in individual stocks that are highly correlated with the funds.

But more generally, and to what might be more useful for those of us trading, what is the best technique?

If individuals are to consider historical performance at all, there are several considerations. The time frames to use -- what lookback period and what holding period? What performance metrics to value -- good recent performance, or poor recent performance? What set of rules -- how to identify patterns that precede profitable trades / switches?

Best, Howard

 

[DeepState]

Greetings --

Deepstate is stating my recommendation correctly. To reiterate the process:

1. Use the analysis of the price history of a list of potentially tradable issues to determine the risk, profit potential, and liquidity of each. Issues that pass those filters become candidates to be traded. They might be individual stocks, individual futures contracts, ETFs that are made up of some group of issues. This is the activity of the "data prospector." No rules (no model) have yet been applied. Independent of any possible rules, some data series are untradable due to high or irregular volatility, others are not worth trading due to low volatility. The individual issues can be given a score to rank them.

2. One at a time, pick a trading candidate and develop a model (a set of rules that identifies probably profitable trades and issues orders to buy and to sell) that trades that single issue long and flat. Or short and flat, but not both in a single model. For stocks and equity ETFs, identifying good signals to be short is more difficult than to be long. Patterns preceding price rises are different than patterns preceding price falls. Do not ask your model to try to identify both and keep them separate.

The process of system development becomes the process of fitting a model to a single data series. It is during this process that we discover whether there are identifiable and persistent signals in the data that precede profitable trades. Not every candidate data series that the prospector gave a passing grade to results in a tradable system.

Use the scientific method. Examine historical data looking for patterns and fit the model to the data, giving the best in-sample fit. Test the model on data not previously used in fitting (validation). The test produces a set of trades that are the best estimate of future performance. Ignore in-sample trade results -- they have no value in estimating future performance. Using the best estimate set of trades, compute metrics giving the maximum safe position size (safe-f) and estimate of compound annual rate of return (CAR25). If this one-time validation shows satisfactory results -- that is a value of CAR25 that is higher than risk free use of the funds -- move the system to trading where it will be managed by a trading management model.

Develop as many such systems as you wish. Each has its own safe-f and CAR25.

3. The trading management model uses trades as its input and has safe-f and CAR25 as its output. The model is based on Bayesian analysis and sequential learning -- learn from the most recent data, adjusting your view as new data is received. Begin with the best estimate set of trades for any system under consideration. Trade, or paper trade, that system producing real trades if real money was used or shadow trades if paper traded. Add the new trades to the best estimate set, reducing the weighting of older trades as new trades are added. Safe-f and CAR25 will change. The change represents changes in the data series. We cannot know whether these are random and within the distribution we began with, or "regime change" coming from a new and different distribution. We act as if they are true indications of distribution shift in the signals coming from the data. Recompute safe-f and CAR25 at every evaluation -- probably at the close of trading every day -- and sort the results. The system to trade next is the one at the top of the list. That is, trade the one system that has the highest risk-normalized CAR. Take a position guided by safe-f. To trade at any position size higher than safe-f risks drawdown that exceeds your tolerance.

Diversity comes from having multiple systems, each of which trades a single issue, available for trading at the next opportunity. As conditions change, risk changes, safe-f changes, CAR25 changes, leadership of the list changes, and the system that is being traded changes -- or, at least, could change.

The portfolio is a portfolio of systems, each trading a single issue with its own set of rules -- rather than a portfolio of issues traded by a commonly applied single set of rules. In terms of modern portfolio theory and the efficient frontier, the chart is one dimensional -- CAR25 along a line -- rather than two dimensional -- risk (typically standard deviation) on one axis, return (typically CAR) on the other. The method I recommend has normalized the risk, so risk is the same for all system. The other dimension is return and that is CAR25.

Best regards, Howard

Thanks for the detailed explanation Howard. It's good to see your thinking laid out here so clearly.

FWIW, I believe that this is actually a trend signal in disguise if the fitting periods are anything like a slightly front weighted 3-12 month period. I won't go through this because, as Quant has suggested, anyone who wants this can do their own work. Not having seen any you have produced, I am confident that this method would work in backtests and across different markets if applied to stocks in broad universes, but less so for diversified ETFs. It would work for commodities and FX. I agree with separating long vs short signals in this case, for stocks but not others.

If producing a great long term risk adjusted return is the aim, the filters to a single stock holding at any given time require an extreme process which essentially assumes that only the 'best' signal of the day has any validity. The underlying belief systems to support that are not reasonable in my view.

The process also makes no allowance for frictions which, depending on the portfolio of systems and their characteristics interacting with the universe of instruments, can destroy an absolutely enormous amoutn of value. For example, a universe of 200 stocks and an infinite number of rules will spit out one which tells you to buy something. The chances of that same something being in place the next day are very low. Let's say we trade once in 2 days. Brokerage of 0.1% each way. That's 126 round trips. 20+% per annum erosion on a fully invested portfolio (yes, you have a safe-f and this changes, but the idea is here). And that is if you take zero spread or slippage.

 

[DeepState]

Greetings --

Deepstate is stating my recommendation correctly. To reiterate the process:

...metrics giving the maximum safe position size (safe-f)...

I really like this aspect of your process which calibrates nominal exposure to risk tolerance and has the additional/associated benefit of volatility stabilisation if the estimation method is valid.

 

[howardbandy]

I hear and understand the attractiveness of long holding periods, low system maintenance, low transaction costs, lower tax rate. Fund managers such as John Bogle (Vanguard funds) and the like recommend this technique -- but they have vested interest in keeping funds in positions to ease their management.

My caution is that the period from 1945 to 2000 is probably far better for long term holding in terms of high reward and low risk than we can expect in the future. If that is repeated, we want to participate. If the future is different -- higher volatility, more efficient markets, significant markets drops -- then we need to have procedures that indicate that risk is increasing and, if possible, ways to profit from changed conditions.

Risk normalization and dynamic position sizing is one such technique. To illustrate, begin with a list of the trades from a series of long term holdings. Note the intra-trade mark-to-market daily high and low equity. Compute the risk of that one equity curve. Form the distribution based on that set of trades and compute the probability of drawdowns of various levels. You will find that risk to the funds exposed to the market is much higher for long term holding. In order to hold risk of the trading account to a reasonable level, some funds must be kept in reserve.

Whatever risk the system has Will Be experienced by the funds exposed to the market. We cannot avoid that. Holding some cash in reserve is what keeps the risk to the trading account reasonable.

But, the funds held in cash earn very little, so overall rate of return is lowered.

Then compare with more active trading -- accurate trades with short holding periods. Normalized for risk, more active trading is safer and more profitable.

Best, Howard

 

[DeepState]

accurate trades with short holding periods.

Is the application of CAR25 the alchemist required to convert mostly noise into accurate signal? With an accurate signal, net of frictions, anything is possible.

Certainly a weaker outlook for equity returns (which seems to be continually defied, just like the edict that bond yields surely can't go negative), lowers the bar for alpha strategies. However, the bar is still there and remains a challenge that 90% will not make in the next 3 years. Add a frictions hurdle of 20%+ on deployed funds and you have a definition of an epic hurdle for a signal to overcome just to break even. Cash could easily turn out to have been the better strategy...or being a broker.

Which still leaves the question on portfolio diversity at any given time (as opposed to through time) somewhat unanswered. Unless the candidate signals are so wild that they switch in and out of predictive ability such that only one exists at any given time and that CAR25 is capable of finding it with some accuracy, this approach is not deploying risk and taking in to account the skill in selecting candidate strategies being filtered for maximum effect - where I am assuming skill exists for the sake of argument. If skill does not exist, as mentioned earlier, it doesn't matter except for saving frictions...so you glide in to the ground more gently than would be the case applying a portfolio approach under the assumption that skill existed in the first place at any level.

 

[howardbandy]

Hi Deepstate --

To my thought, the procedure is pretty clear and pretty much all trading system developers have the requisite domain knowledge.
... Development:
1. Use the data prospector to curate a list of potential candidates.
2. For each candidate, attempt to develop a long / flat trading system using the scientific method. The model is rule-based, uses state signals, marks-to-market daily, is looking for trades that are accurate, frequent, and hold a day or two. If a traditional trading system development platform is being used, an oscillator with a fast cycle works. Pick any one, since at that rate, all are equivalent. Or some other indicator of your choice that will give about 50 signals per year.
3. Also attempt to develop short / flat systems, but there will be lower success.
... Trading:
1. Bring data up to date and evaluate all the systems daily, computing safe-f and CAR25. If the one with the highest CAR25 has a score higher than a risk free account, take a position using safe-f portion of the account.
2. As the jingle says -- rinse and repeat daily.

Still one system at a time, one day at a time.

Volatility was one of the characteristics examined by the prospector. When he or she passed it originally, volatility was acceptable. If it increases at a later time, dynamic position sizing will identify that and lower safe-f and the CAR25 score, moving it down the list of those available to be traded.

If AmiBroker is being used, I have published several models that work.

If Python / machine learning is being used, I have published one decision tree model showing that it replicates the traditional platform. Also examples of about ten alternative models that can be applied with little change to anything. And suggestions for additional predictive variables.

Modeling experience and skill become more important here. The magic sauce, whose recipe developers of machine learning trading systems will hold close, is which additional predictor variables are useful, what steps are taken to create stationary data, and what data preparation is used.

Best regards, Howard

 

[ducati916]

My caution is that the period from 1945 to 2000 is probably far better for long term holding in terms of high reward and low risk than we can expect in the future. If that is repeated, we want to participate. If the future is different -- higher volatility, more efficient markets, significant markets drops -- then we need to have procedures that indicate that risk is increasing and, if possible, ways to profit from changed conditions.

In that time frame there have been two periods [in the US] of extended high volatility and sideways movement. The first 1965 - 1985 and the second 1999 - 2014.

That is one period of 20 years and the one period of 15 years, within your 72 year time frame, which is more or less, 50% of the time.

If this period is considered a good period with regards to risk/reward and the future could contain higher volatility, which changes risk/reward for the worse, then 'we need to have procedures that indicate that risk is increasing and, if possible, ways to profit from changed conditions'

It would seem that you advocate more frequent trading [shorter signals]. Correct me if this is not the case.

I would argue that for many retail traders, this is simply not the way forward. Intra-day and daily price moves are simply too noisy.

jog on
duc