Model discussion

[Pierre Brodeur]

(2) Make the threshold choice such that performance isn't overly sensitive to slight changes of threshold value

The problem with OS threshold misses is that they inevitably lead to a large loss in the form of a string of mis-directed trades (repeated attempts to re-short an uptrend instead of being in buy mode). Only testing can check whether requirement #2 above holds with a choice of -70.
Trader D

No indicator is perfect, of course and as traders we have the luxury of being able to put any indicator in the current market dynamics context which many model have difficulty doing. That is why many modern models (especially risk models) have volatility regime adjustments to calibrate factor volatilities to current market levels. But that is a very difficult thing to do if only because risk is non linear. Others use nonstationary stochastic processes (ARCH & GARCH models) in order to deal with the fact that model coefficient do change over time. However, RSI for example has the advantage of being between 0% and 100% and thus if this would reflect the distribution ( normal or otherwise) of historical value (OB/OS) it would be a major improvement over the absolute indicators currently available or the current arbitrary OB/OS hard coded numbers (i.e.: 70) currently being used by Pascal.

My gut feeling is that this could be a problem without use of a more relaxed direction determinant, possibly involving another independent indicator.
Trader D

I believe we are in agreement on the need for another independent indicator.

[mingpan.lam]

Here's my opinions

1. 20DMF
--> agreed with Timothy, increased the numbers of ETF the robots trade, for example, when most ETF is in buy mode, the remaining neutral mode ETF cannot take short term short entry (a failed safe mechanism), alternatively, when most ETF is in short mode, the remaining neutral mode ETF cannot take short term long entry


2. The GDX MF

Model Weaknesses

• The model acts very fast on a signal change but might be prone to whipsaws, mostly due to the underlying volatility.
--> it happened in Dec last year, does it really due to volatility or does it due to the low volume because of holiday? Or, volatility due to the low volume?


3. The Robots

• Update the statistical trading tables for both robots
--> is it possible to automate the process for the statistical trading tables?


4. RT On-going development work

--> sms alert (can be done via twitter)

5. Sector Rotation (SR) trading model

--> can we use GDX MF model to run on each different sectors? Then we don't need to worry about the stock.

--> I think real time system will be v. useful to provide a better entry and tight stop but we need an real time alert system.

[senco]

The biggest challenge we have is a very short history of data available for backtesting; any idea that will resolve the few occurrences where the model broke down cannot be confirmed with statistical confidence. For this reason, the 20DMF had a couple of tweaks in the past. Were those optimal? We shall probably know only in many years from now (n fact, we trust the model because it makes fundamental sense, not because of a thorough out of sample statistical validation – for this we do not have sufficient data points. Adaptive OB/OS determination makes sense and it might be more robust than an arbitrary level. Maybe.
So what can one do? On a conceptual level: (a) keep the model as simple as possible – the less parameters, decisions and ‘knobs’ there are - the less brittle it will be, (b) introduce additional data points of a somewhat different ilk by incorporating other indicators (for decision, confirmation, or vote), and (c) use several systems/robots for diversification. Well, all ideas mentioned earlier in this thread :-)
On a practical level – there are a number of ‘breadth related’ indicators that could be used quite effectively to (i) identify bottoms - maybe combine with 20DMF in some voting mechanism, and (ii) identify a bullish state of the market – to get the 20DMF out of a ‘neutral’ state, and/or be used together either in a voting, or an allocation mechanism.
By ‘breadth related’ indicators I mean things like: new highs/new lows, volume or issues advance/decline, TICK, TRIN (Arms), number of issues over or crossing a moving average. These days the data for these indicators can be easily accessed in real time - see my comment in the Tradestation thread.
Breadth models try to measure the underlying happenings in the market as the MF does, though in a different way - and they can be used to create good timing models on their own. There is a good possibility that combining them with 20DMF will increase the model's robustness.

[senco]

While at it, a couple of general thoughts:

- Whipsaws: If the losses on whipsaws are small, best to look at those just as the cost of doing business. Many systems whipsawed in the huge volatility of last year much more than in the last twenty years, simply bringing out the reality of the market... politicians and central banks flip-flopping.

- When modifying / tweaking a model, it is a good idea to continue and maintain full data series (past and future) of both versions, not just discard the old one. In my experience one can still learn from systems abandoned many years ago.

- We are very interested in the Maximum Drawdown of a backtest (and more than that, of a system we trade live). It is also important to understand that the MDD does not represent well the statistics of a trading system output (it is the outcome of a specific path in time, out of many that could have happened). Therefore, MDD is not a good predictor of a system’s future drawdown and is not a good measure for a system’s risk. The saying “your worst drawdown did not happen yet” has indeed a theoretical basis. When comparing different versions of a system in development - it is much better to use measures with more statistical contents, like a rolling period downward deviation.

[TraderD]

While at it, a couple of general thoughts:
- Whipsaws: If the losses on whipsaws are small, best to look at those just as the cost of doing business. Many systems whipsawed in the huge volatility of last year much more than in the last twenty years, simply bringing out the reality of the market... politicians and central banks flip-flopping.

AFAICT, EV as a follower of large money can only be as good as they are. 2011 performance of large funds have been rather dismal (an understatement), so it's likely to expect EV to deteriorate in its predictive ability. Paul in a recent GGT post shows a bunch of examples where contrary to LEV divergence, price keeps going in the other direction, which begs the use of this technique with other indicator(s) as you've suggested, or at least a much more accurate understanding where EV does work.

- We are very interested in the Maximum Drawdown of a backtest (and more than that, of a system we trade live). It is also important to understand that the MDD does not represent well the statistics of a trading system output (it is the outcome of a specific path in time, out of many that could have happened). Therefore, MDD is not a good predictor of a system’s future drawdown and is not a good measure for a system’s risk. The saying “your worst drawdown did not happen yet” has indeed a theoretical basis. When comparing different versions of a system in development - it is much better to use measures with more statistical contents, like a rolling period downward deviation.

Fully agree with MDD comment. Do you have a link to RPDD stat calculations?

[Pascal]

AFAICT, EV as a follower of large money can only be as good as they are. 2011 performance of large funds have been rather dismal (an understatement), so it's likely to expect EV to deteriorate in its predictive ability. Paul in a recent GGT post shows a bunch of examples where contrary to LEV divergence, price keeps going in the other direction, which begs the use of this technique with other indicator(s) as you've suggested, or at least a much more accurate understanding where EV does work.

Fully agree with MDD comment. Do you have a link to RPDD stat calculations?

EV detects an equilibrium and not a force. It statistically detects when money comes in/out. For a single stock, the movements of money will often be opposite to the price moves, because large players use available liquidity to buy/sell. However, when the price is in a trading range or at a turning point, EV will often show what is happening below the surface and what the next move will be. This is why I always use AB/LER as a combination.

The money flow is however more predictive when it is collected by sectors or industry group.

Also, large money does not necessarily mean that these are large funds. It could be some artificial FED liquidity injection. Therefore, I believe that your statement that EV's predictive ability will deteriorate might or might not be true. I however believe that EV still gives early warnings of what large money is doing. A few years ago, I noted that it could give up to one day advance warning. Today, because the level field is pretty high and everybody is running fast computers, the warning time is probably less than a day.

Anyway, I still prefer to know where the money is going than not to know it.


Pascal

[Pascal]

I am attaching here a general figure of the different models that are in use, but also of a back-test campaign made last week using these models.

We can see that the one level that cannot be traded is the sectors level, simply because these are sectors that I have defined myself and for which there is no instrument. These sectors are mainly used either for the 20DMF or for the stock filters.



Below are the yearly returns of the S&P and the 20DMF. These are returns compounded within one year. If a 20DMF trade is overlapping two years, I separated taking one part in one year and the second in the following year. This way, we can have a better comparison.



The objective of this back-test work was to measure the relevance to trade sector information.
Is it good to buy a sector when the sector issues a buy signal and short it when it issues a short signal?

The results are below. We can see that indeed, sectors trading is better than B/H on the S&P 500.
However, sectors trading in sync with the 20DMF is still better. Unfortunately, "in Sync" does not give better results than the 20DMF itself.

These results are not surprising: they are "in line" with what I had when I did a similar back-test two years ago.
My conclusion at that time was that even when a sector is flashing a buy signal, it still needs to be as close as possible from a 20DMF buy signal. The later we are from that signal, the worst the returns.



Therefore, the next idea was to select only five sectors when the 20DMF issued a signal and then from each of these sectors, get all the stocks AB/LER data and select the five stocks that were showing the best AB/LER combination. The results of such a test are shown below. These are also in line with the results I had two years ago. The table below show us that whatever efforts we can do to select specific stocks, it will be hard to beat a two time leveraged ETF that trades the 20DMF signals. Of course, specific stock trading might lead to lower DD (I did not calculate such DD.)



I however did another test that gives interesting results.
I selected one date in the past and decided to trade either long or short by intervals of 20 days from that date.
this means at day one, I buy the five best sectors and I sell them at day 20.
The five best sectors are those that show the weakest price RS.
On day 20, I again select the five best sectors and buy them.

On exactly the same day, I also short the five sectors that are the most overbought and I sell these positions 20 days later.

The results are shown in the table below.

We can see that this dumb strategy worked well for longs in 2009 and for shorts in 2008.
In 2010 and 2011, it did not work that well.

However, this strategy shows something important: the rotational aspect of the market. It shows that it makes sense to rotate money. This is of course obvious! I still prefer to see it in the data than not seeing it, because this means that it will probably make sense to develop a set of ETFs MF by industry groups and rotate between them.




I will now start working on these industry group MF models.


Pascal

[Pascal]

While at it, a couple of general thoughts:

- Whipsaws: If the losses on whipsaws are small, best to look at those just as the cost of doing business. Many systems whipsawed in the huge volatility of last year much more than in the last twenty years, simply bringing out the reality of the market... politicians and central banks flip-flopping.

Whipsaws is a clear limitation in the EV based models and must be dealt with with great care.
Indeed, we have seen the EV often moves in a direction opposite to price on single stocks, because large players would take advantage of higher liquidity to buy/sell positions contrary to the price move. However, when a stock/Sector is on its lower boundary or in oversold, then a bounce in EV might indicate that there is real accumulation because the stock/sector is "cheap".

Hence, when I applied the OB/OS MF model to the 96 sectors that I defined, I noticed that the return was lower than using the usual simpler sector model to buy and short. The reason was simply that the sectors are in a "hectic" manner when in OB/OS. They will switch up/down until they stabilize and move definitively in a new direction. I believe that this is a factor that is "inherent" to EV , especially on single stocks or on a basket of a few stocks.

However, industry groups and total market level measures are less prone to EV whipsaws, because they do compensate each other. We would need the majority of the stocks in the sector to be bought and then sold the next day to have whipsaws on an industry level. This has fewer chances to occur and hence, OB/OS works better the more stocks you use in the basket.


Pascal

[xr-3609]

WOW! Great work! It looks like we need a new instrument to trade the EV Sectors......maybe a "synthetic EV ETF" based on the best five sectors/stocks when a 20DMF signal comes down (or up)!

Greg

[TraderD]

No indicator is perfect, of course and as traders we have the luxury of being able to put any indicator in the current market dynamics context which many model have difficulty doing. That is why many modern models (especially risk models) have volatility regime adjustments to calibrate factor volatilities to current market levels. But that is a very difficult thing to do if only because risk is non linear. Others use nonstationary stochastic processes (ARCH & GARCH models) in order to deal with the fact that model coefficient do change over time. However, RSI for example has the advantage of being between 0% and 100% and thus if this would reflect the distribution ( normal or otherwise) of historical value (OB/OS) it would be a major improvement over the absolute indicators currently available or the current arbitrary OB/OS hard coded numbers (i.e.: 70) currently being used by Pascal.

IIUC, portfolio factor calibration, unlike trading with stops, is prone to occasional black swans where the fat tail of the distribution isn't properly captured by the model. Add leverage to that with a quest for market-beating returns and you got a prescription for a blow up. I'm not sure I see the benefit of RSI over the current OB/OS oscillator, which is also bounded by a fixed range (-100,+100). Is it the non-linear mapping? RSI would also need to utilize threshold choices (typically 30/70).

Trader D