Does Any Technical Analysis Work? (Part 6)

Today I conclude with my internet sample of opinions on whether technical analysis (TA) can actually work.

Here’s an excerpt from an article by Proinsias O’Mahony:

     > While TA remains widely used, that doesn’t mean it’s not bunkum.
     > Indeed, many technical traders would be the first to accept that
     > the field is full of charlatans. As bond expert and author Martin
     > Fridson has written: “The only thing we know for certain about
     > TA is that it’s possible to make a living publishing a newsletter
     > on the subject.”

See my post here.

     > Such newsletters are full of references to obscure Japanese
     > candlestick chart patterns, Elliott Wave theory, Fibonacci
     > numbers, and all kinds of other vague and unverified assertions.

I wrote about some of this here.

After all the opinions in these five posts, what do we have?

The four excerpts from Part 2 are optimistic and give me hope.

The two excerpts from Part 3 suggest that anything we find in books, seminars, or on the internet will not work.

The commentary from Part 4 suggests that no simple indicators will work and that bots available for purchase will not work over the long term. The latter is not new because I don’t expect any system to work forever.

[9] – [11] in Part 5 argue that no bots for sale will ever work, no .pdf guidelines or strategies based in TA will work, no simple TA-based systems will work, and that institutions advertising advanced TA-based departments with oodles of computing power and academic doctorates are really for marketing purposes only.

Today’s excerpts argue against TA like most of the others I have surveyed.

Of all this, I have to hope [1] – [4] is correct. It may not be easy, though. Although my initial impressions do not concur, at least one prolific writer in the algorithmic trading space says simple strategies are all we need.

As an aside, O’Mahony also provides some ammunition in case I ever want to sound off on fundamentals again (see second-to-last paragraph here):

     > The same can be said of fundamental analysis, however.
     > Billionaire investor Mark Cuban once scoffed that “fundamentals
     > is a word invented by sellers to find buyers… metrics created
     > to help stockbrokers sell stocks, and to give buyers reassurance
     > when buying stocks.”
     >
     > That may be cynical, but it’s a documented fact that most
     > fundamental managers underperform the markets. Most investors
     > would be better adopting a buy-and-hold approach rather than
     > painstakingly studying stockbroker notes in a futile attempt
     > to gain an edge.

Do you know if any successful traders even exist?

Next time, I will come back to discussion of my personal experience thus far.

Does Any Technical Analysis Work? (Part 4)

One can do an internet search on “does technical analysis work” and find lots of discussion on the topic.

JA wrote:

     > …I know if someone discovers a profitable strategy he won’t share
     > it… I was asking about “public and very common strategies using
     > TA,” like RSI, Fibonacci etc, because if they work, why aren’t a
     > lot of bots out there making money really easy with that? And my
     > comment about people selling courses is because it confuses me
     > when I see people I trust telling me they use those techniques
     > with good results. How is that possible? Do you think they work?

CF responded:

     > ok dude. you know the “public and very common strategies like
     > RSI, Fibonacci etc.” Those are absolutely total trash. Like: a joke,
     > just a total joke. It’s quite confusing to me that you seem to be
     > saying “many people think those work…” If you ask 1000 working
     > daytraders… they will just smile. Those things are equivalent to
     > the “lose weight!” systems you see sold on informercials you know?
     >
     > With regard to people telling you they use those techniques with
     > good results, those people are: liars. Even if it’s your brother
     > or something, non-working traders often get really confused, sort
     > of exaggerating their own trade results for the better! [7]

Not to be lost on the humor (remember back to [6]! The internet parlance “ROFL” comes to mind right now) is this claim that the standard indicators commonly discussed as the basics of TA do not work.

AJ writes:

     > …any developer can program a bot to trade using TA… and there
     > are a lot of bots out there, and none of them are working. I did
     > a lot of backtesting myself using multiple strategies and none
     > of them work long term… do you really think there is some bot
     > out there that can make a good profit through automated trading
     > and is actually being sold?? Yeah sure, so I coded a bot to make
     > millions when I’m sleeping and I’m publishing it as open source
     > or selling it for $500. Just name ONE that makes guaranteed
     > profit. [8]

He absolutely may be right about this!

On another note, I’m not at all convinced that he knows “none of them are working.” I’m not convinced he’s ever purchased/leased a single bot to try for himself. I’m not convinced he has had real conversation with [credible] others who have. He may be one to spend lots of time in the forums and read what others (also potentially with limited experience like himself) have to say. Back from [7], does anybody know 1,000 working daytraders?” Lots of people say lots of things. We have to evaluate the source, too, and assess what we really don’t (and what they can’t) know.

Again, AJ makes some really good points to consider. As mentioned in the fifth paragraph here, I’ve just started testing multiple strategies over the long term. None have really worked thus far.

To be continued next time.

Does Any Technical Analysis Work? (Part 3)

Today I want to continue discussing some other internet opinions I’ve found about whether TA can actually work.

     > There are a lot of people giving signals to buy or sell based on some
     > indicator or price pattern. There are a lot of people selling courses
     > about how to make profit trading using these well known strategies.
     > And I don’t mean a really… [obscure and esoteric genius strategy]…
     > I mean all these Fib retracements, RSI, MA, etc. I can’t understand why
     > those strategies exist. They are public and they are easy to implement
     > with a bot, so they cannot work right?
     >
     > No you’re not missing anything, if it worked it wouldn’t be on sale.
     > But that’s not a wholesale indictment of TA. [5]

Lots and lots of empty claims exist out there with very little supporting data. All this is to say we need to generate the data ourselves. We need to test everything else run the risk of being duped by something that sounds good and is marketed well.

A “bot,” by the way, is a computer executing algorithmic trading code.

On one forum, I saw someone repeat my observation given in the second paragraph of the above-linked post. Here is another’s response to that:

     > You must be joking man, there are a lot of people selling books on
     > how to lose weight! Good grief! [6]

Made me laugh!

Not only is this a great analogy, though, I suspect it may be true. Judging by how many diets have crossed the screens and shelves over the years, how many can really be said to work? Different diets work for different people here and there, but most diets probably have a low success rate overall. The same is probably true for TA because emotionless systems are very hard to maintain through times of famine and decline. Unless someone has done thorough trading system development, fear will probably own the day forcing said trader to sit alone on the sidelines licking psychological wounds from locked-in losses.

I will continue next time.

Curve Fitting 101 (Part 4)

Before I move forward with trading system development approaches, I want to give one more example of how to curve fit: increasing strategy complexity.

Here’s a basic strategy (in EasyLanguage):

          Input: len(14), //optimize 2-20 by 3
                    offset(40), //optimize 30-45 by 5
                    pt(500); //optimize 500-2000 by 500

          setprofittarget(pt);

          if percentR(len) < (50 - offset) then sellshort next bar at market;
          if percentR(len) > (50 + offset) then buy next bar at market;

This is a trend-following strategy that will go long (short) when Williams %R is above (below) an extreme value. I am testing extreme %R values of 20 (80), 15 (85), 10 (90), and 5 (95):

Notice how the strategy performed exceedingly well until the arrow (~500 days in). After that, equity took a precipitous fall and lingered in the doldrums for the remainder of the backtest.

I can improve this by increasing strategy complexity. At the arrow, I will flip from trend following to mean reversion:

          if percentR(len) < (50 - offset) then buy next bar at market;
          if percentR(len) > (50 + offset) then sellshort next bar at market;

After adding a rule, I can look at the equity curve again. Where I see the beginning of another large drawdown, I will add a subsequent rule to flip back to trend following. I can then look at the equity curve again. Where I see the beginning of another large drawdown, I will add a subsequent rule to flip to mean reversion. Rinse and repeat a few times!

With first four lines of code unchanged, here’s what I ended up with:

          If date <= ELDate(11,01,2011) then begin
               if percentR(len) < (50 - offset) then sellshort next bar at market;
               if percentR(len) > (50 + offset) then buy next bar at market;
          end;

          If date > ELDate(11,01,2011) and date < ELDate(08,14,2012) then begin
               if percentR(len) < (50 - offset) then buy next bar at market;
               if percentR(len) > (50 + offset) then sellshort next bar at market;
          end;

          If date >= ELDate(08,14,2012) and date < ELDate(02,20,2013) then begin
               if percentR(len) < (50 - offset) then sellshort next bar at market;
               if percentR(len) > (50 + offset) then buy next bar at market;
          end;

          If date >= ELDate(02,20,2013) and date < ELDate(02,11,2014) then begin
               if percentR(len) < (50 - offset) then buy next bar at market;
               if percentR(len) > (50 + offset) then sellshort next bar at market;
          end;

          If date >= ELDate(02,11,2014) and date < ELDate(12,03,2015) then begin
               if percentR(len) < (50 - offset) then sellshort next bar at market;
               if percentR(len) > (50 + offset) then buy next bar at market;
          end;

          If date >= ELDate(12,03,2015) and date < ELDate(09,29,2016) then begin
               if percentR(len) < (50 - offset) then buy next bar at market;
               if percentR(len) > (50 + offset) then sellshort next bar at market;
          end;

          If date >= ELDate(09,29,2016) and date < ELDate(01,24,2018) then begin
               if percentR(len) < (50 - offset) then sellshort next bar at market;
               if percentR(len) > (50 + offset) then buy next bar at market;
          end;

          If date >= ELDate(01,24,2018) then begin
               if percentR(len) < (50 - offset) then buy next bar at market;
               if percentR(len) > (50 + offset) then sellshort next bar at market;
          end;

I increased the number of rules in this strategy from one to eight. The rules couple date ranges with a mean-reversion/trend-following switch.

This is some of the most blatant curve-fitting possible. I’ve used nothing but hindsight, which we know is invalid as a system development technique (see fourth paragraph). Also, recall in Part 2 where I chose a specific parameter value (VIX < 35)? With that approach, I still have to settle for whatever equity curve results. Here, every rule I add directly manipulates the equity curve in whatever direction I so choose.

This is also the epitome of curve fitting because the rules refer to historical date ranges that don’t even exist in the future. Doing so is criminal indeed:

This has 78% winners on 479 trades, a profit factor of 1.71, and trades nearly even with regard to long and short positions (with short trade profit factor of 1.55). Slippage and commissions are included.

The more complicated the strategy, the more beautiful the equity curves. Just remember the more beautiful the equity curve, the less likely it is to be fit to signal rather than noise.

Today’s result is totally, utterly, and absolutely curve-fit. Can’t trade that, folks: no way, no how.

Curve Fitting 101 (Part 2)

In the last post, I gave examples of how optimization could be used to manufacturer an investment newsletter scam. Today, I want to get back to trading system development and give another example of how curve fitting can generate deceptive results.

Also known as overfitting, curve fitting is tailoring a trading strategy to better match historical noise than persistent signal. The strategy may look great in backtesting, but it is unlikely to look good when applied to future data because noise is random.

Besides optimization (see Part 1), a second way to curve fit is to use specific parameter values. Consider a simple strategy that buys pullbacks (in TradeStation’s EasyLanguage):

          If C = lowest(C,10) then buy next bar at market;
          If barssinceentry >= 3 then sell next bar at market;

This strategy buys after the lowest close in 10 days (bars) and sells after the position has been open for at least three days.

Suppose this strategy suffered its maximum drawdown on the VIX spike over 35 in the 2008-9 financial crisis. Let’s add another rule by changing the first line to:

           If C = lowest(C,10) and close of data2 < 35 then buy next bar at market;

Now, the long entry will occur when close is the lowest close in 10 days and when VIX is less than 35. This can improve the equity curve by avoiding a losing trade(s) and going flat in an area that previously experienced drawdown.

The problem here is adding a specific rule to fix a particular past occurrence. It’s like applying a single Band-Aid. A similar event capable of causing significant damage may occur in the future, but the filter may not be effective since we based it on a sample size of 1. Look what happened in February 2018:

The arrows point to February 2 when ES had a big drop of 2.3% and VIX spiked to 17.31. That was a big move with VIX closing at its highest level in nearly 15 months (11/8/16), but it did not trigger our filter (17.31 < 35). The curve-fit strategy would enter the pullback just in time for February 5, 2018, when the market fell 5.4% and VIX spiked to 37.32. The VIX filter would now be active: one bar too late to serve its purpose. I analyzed a trading strategy that applied this rule here. My curve-fitting comments are in the seventh paragraph.

I will continue the discussion next time.