In defense of a quantitative attack to financialdannbsp;markets

I have the feeling that in that respect is few subtle yet banquet misconception about data-driven research in financial markets and I testament bring down this article: Quest Alpha – Not Even off Untimely: Wherefore Data-Well-mined Market Predictions Are Worse Than Uselessdannbsp;by Justice Litle (also coming into court in his website:dannbsp;Mercenary Trader) equally a starting indicate for the discussion.

The article itselfdannbsp;is born as a rant against this article connected Yahoo Finance:dannbsp;Why Boring Is Bullish, which "infers" a 89% chance of bullish action on the Sdanamp;P supported on a sample of 18 previous cases where wedannbsp;had similar "low vol" as now.
Now, let me clearly say that the Yahoodannbsp;article is indefensible for a number of reasons in my opinion (to mention a couple of: style too small sample size, no robustness psychoanalysis, no mention of numbers of trials that were run), so in this I agree with Mr Litle.

However Mr Litle goes on the far side this anddannbsp;explains wherefore equity markets cannot be "irksome" right at once:

"The potential flight of equity markets is DIRECTLY IMPACTED BY the trajectory of debt and currency markets (which are the OPPOSITE of boring now). […] "Calm ahead the storm boring," maybe. Plain old ho-hum boring? Ah, no." […]

and eventually moves his criticsdannbsp;to information mining in financial markets in general:

"Markets are far from simple. In fact they are real complex. In and of itself, predictions settled on data mining of a single historical inconstant or single cherry-picked pattern observation are well-nig e'er worse than useless because they ignore a burden merging of factors." […]

"When information technology comes to predicting approaching outputs of colonial systems, virtually ALL forms of single-variable statistical thinking are flawed."dannbsp;[…]

"The sole way to avoid getting fooled by spurious data or superficial thinking is to put real elbow grease into truly understanding what drives markets and why…and at one time you have that understanding you don't need to cherry pick or data mine because you have something better: The power to assess a confluence of primal factors in the on hand, equally impacting important market relationships Here and now."

Now, while I agree that financial markets are very complex and that IT's verysimplified to atomic number 4 fooled, I believe that these statements just about data mining are a shade too generic.
Using a single historical variable or taking into thoughtfulness the charm of multiple factors says absolutely nothing per sedannbsp;on how bang-up a prediction is (anddannbsp;with "prediction" I refer to any kind of statistical inference over thedannbsp;rising).

Generally, to cost healthy to make a prediction with severaldannbsp;valuedannbsp;same has to identify certain features (variables) that concerted in a bound way have several predictivedannbsp;power concluded future events. This is true for any field and for any prevision method, glucinium information technology AI surgery human reasoning.
The heavy percentage of course is finding these features and combining them.

Looking at things this way, the author of the Yahoodannbsp;article is just claimingdannbsp;that (a certain definition of) low-degree of volatility has some informative index overdannbsp;later returns. What Mr Litle is respondingdannbsp;is that monetary policy, debt and currency markets or else are better features to enjoyment, supported his experience and view of the world.

Is this genuinely that different from properly cooked information mining?

The double question is whether "discernment" the causes of certain market dynamics is a key factor in making them forecastable to a sure as shooting degree(note of hand the quotes in "understanding").

I don't believe this to be the case.

To pull in adannbsp;parallel with the world of Physics, physicists certainly don't always understand WHY certain things follow a predictable law. Rather they observe a certain behaviour and theydannbsp;try to describe it. If on the way they can happen some kinda explanation for it, the better. But in that respect will always be an additional "why" which requires an answer (why apples fall towards the ground? -dangt; gravity -dangt; why solemnity exist? -dangt; relativity -dangt; etc).

Of course adannbsp;key difference withdannbsp;Physicsdannbsp;is that commercial enterprise markets cannot be only described aside equations, beingness the results of complex interactions of billions of masses. From a practical repoint of view this means that with a data-driven approach we have to put much more attention in developing a framework to evaluate the actual prophetical power of any model, which too will scarcely work "forever".

But correspondingdannbsp;difficulties apply todannbsp;any kind of discretionary trading.dannbsp;The very same fact that there are so many factors in play (and hence so some noise) makes it hard for our brainiac to analyze the situation objectively, and surely the manydannbsp;psychological feature biasesdannbsp;that regard United States don't aid.
So ourdannbsp;"understanding" of the causes of market movementsdannbsp;can't really go that far.dannbsp;E.g. wedannbsp;might understand that a certain inefficiency exists because of some institutions operating under some constraints, but wedannbsp;won't love how overnight these constraints will stay in place or when some competitors will pick ondannbsp;this inefficiency reduction our margin of profit operating theatre still causing the markets todannbsp;behave in a whole unpredictable way of life.

With this I Don River't want to say that usingdannbsp;much free will is wasted – quite I'm fitting trying to argue that there is a put for some in trading and I seedannbsp;narydannbsp;dualism here.Pure (properly done)dannbsp;data-driven research and fresh big/discretional research lead to two different sets of opportunities that give the sack besides overlap in some situations.

Probably discretionary trading can be more responsive to changing market dynamics, whereas a data-driven approach might have his strength in the portabilitydannbsp;of the operations to contrastive markets and in how quantitative IT is.
And in any case I powerfully believe that any data-impelled depth psychology is only arsenic good as the persuasion we put into IT, and likewise any type of discretionary trading hind end only profit fromdannbsp;devising use of some quantitative analytic thinking.

To comment on a last point brought up by Mr Litle:

"Yet we spend roughly zero time along data mining, with no interest in statements equivalent "Over the past X age, the Sdanamp;P did this X percent of the time."

Why this contrast? Because markets are a complex oversea of moving and interlocking variables – and it is the historical drivers and qualitative causal agency-effect relationships are what have unceasing prise. It is not the output of a spreadsheet that matters – the radiation diagram-based cherry picking lacking insight as to what created the results – only the qualitative relationships truly attributable to marijuana cigarette causing of various outcomes, connected a event-by-case basis, with a very larger nod to chronicle and context."

I agree that what matters is indeed determination close to "relationships" that have real predictive power over the forthcoming. Butdannbsp;how unity finds these relationships is a complex weigh and ane has to dannbsp;dig in the specifics of each case to find outer if the analysis has some value, because generally speech production the output of a spreadsheet can be as good OR as bad as any qualitative relationships one may think todannbsp;hold.

Andrea

Order coordinated algorithms

In now's markets dominated by High-Frequency algos, room for net profit for not-Atomic number 72 (and Sir Thomas More importantly, non-HF aware) guys is generally speaking reduced. The proportional performance touch on of HF is likely to bedannbsp;bigger the smaller is your average barter anddannbsp;the shorter your holding catamenia.

However, in my experience this doesn't have to be necessarydannbsp;the case: plainly put, as in any business you have to adapt to the competitorsdannbsp;and in this vitrine one way of doingdannbsp;it is to pay more attention and improvedannbsp;thedannbsp;capital punishment side of your trading. This is not always easy realizable (see the "Timestamp put-on" reportable by Zerohedge), just there are some low-dangling fruits that can be picked equally a opening.

If thisdannbsp;statement English hawthorn uninjured kind of obscure to you, I have an lesson based on my experience that supports it and that I thinkdannbsp;that could make up useful to others (while hopefully non having too much of an impact on mydannbsp;strategies).

While all my modelsdannbsp;are fully automated, I still like to see at markets and particularly at order books when my orders are being executed.
Something that I noticed quite much prison term ago when trading 30y U.S. bond futures was that whenever my limit orders were executed, I was immediately at a loss.
What this means is better explained by an example. Say that we haddannbsp;an order hold that looked like this:

and that my sell limit order was enclosed indannbsp;those 750 @ 134.6.

Whenever I was dead, the mid-price would then immediatelydannbsp;move against me, and the book would so look something like this:


bid_offer_2

Au fon what was occurrent was that my orderdannbsp;was always ane of the last to atomic number 4 executed, so the simple fact that it got full meant that there were no longer offers (bids) at mydannbsp;level, and the Sunday-go-to-meeting bid and offer would move ahead (down) one ticking.

A nimbledannbsp;probe on the CME web site unconcealed that the cause for this was the type of order matching algo beingness misused by the exchange, a First In, First Out (FIFO) algo.

What is a co-ordinated algorithm?
CME explains it as:

A matching algorithm is a technique to allocate mated quantities, misused whendannbsp;an attacker order matches with i or multiple resting orders. Algorithms apply to both outright and implied twin .

In Rajeev Ranjan's internet sitedannbsp;you can line up a more in-profundity launching to Order Matching Algorithms (as well as other resources on HFT/algo trading).

In the model above, my trading model was instructed to send the fix say only if the price was encompassing enough to my desired plane, which foreverdannbsp;ready-made me unmatchable of the last to join the queue up and thu one of the last to embody filled, according to the FIFO prototype.

In pragmatic terms, what this meant was that I was always executed indannbsp;the bad possible scenarios, that is when the damage would continuedannbsp;in the opposite counseling of my set up, and simultaneously I was never executed in the best scenarios, that is when the price would "touch" my level and then reverse spine in my favour.

As you can opine, a simple workaround for me was to send my determine orders (when operational under First in first out matching algos) arsenic early as practical, simply more often than not speaking, this observation can suggestdannbsp;different things to unusual citizenry.dannbsp;For day traders that are non tradingdannbsp;in an machine-driven fashion, operating underdannbsp;FIFO matching algorithms could oft intend increasing one'sdannbsp;Maximum Harmful Executiondannbsp;by one tick (which candannbsp;follow quite a circle, depending on what one is doing ), unless one isdannbsp;able to playact around information technology.

Likewise to this case, there are early situations when the Order matching algo in use and trades execution in general can become as important as the strategies/trade ideas themselves.

Another example of making good use of order matching algorithms could be that of a trader operational under a pro-ratadannbsp;matching algorithmic program, typical of Eurodollar (IR) futures. If youdannbsp;real want a fill of X lots, you could just send an order that issomewhatdannbsp;bigger than X – with the extra measure beingness set by how aggressive you want/need to be – and erstwhile filled try to cancel the remaining tons (DISCLAIMER: of run over by doing this you are actively risking of being full in all the dozens, sol just get into't take my word on this beingness a good practice and ut it at your own risk).

Of coursedannbsp;paying attention to the matching algorithm is just scratching the surface of the High-Frequency world, but I would think that in roughly situations it's an easy "scratch" to do and one that could straight offdannbsp;hyperkinetic syndrome some rate.

To conclude this mail service, let me clearly say that for how gooddannbsp;ourdannbsp;grocery store simulator is, trades execution can't always beryllium modelled beforehand. This doesn't mean that wedannbsp;should give in the lead trying to wee simulations as true to life (and somewhat conservative) as thinkable, e.g. in terms of fills and slippage (here's a nice post happening what is slippage by Prof. Wash up Balch). Rather, we should just rememberdannbsp;that there is no very substitute for personal low hand observation and interaction with the universe.
On the whole, it shouldn't really come as a surprisal that simple observation is a reigning tool, being IT the opening of the scientific method acting.

Andrea

Feature excerption in tradingdannbsp;algorithms

Of late I have been looking a more systematised way to get around overfitting and in my quest I found it functional to take up some techniques from the Simple machine Eruditeness field.

If you think of information technology, a trading algorithm is just a descriptor of AI applied to prices series.dannbsp;This statement, althoughdannbsp;possibly open-and-shut, puts us in the position to apply a number of Machine Learning techniques to our trading strategies design.

Expanding what discusseddannbsp;heredannbsp;(and here), it seems intuitive that the moredannbsp;featuresdannbsp;in a model, the more generally speakingdannbsp;the model mightiness make up subject to overfitting. This problem is known as the bias-variance trade-offdannbsp;and is usually summarised by the chart happening the right.

bias-variance tradeoff

American Samoa complexity increases, performance in the training set increases while prediction power degrades

What's possibly less intuitive is that the taxonomic group features utilized indannbsp;relation with the kinetics to predict play a key office in determining whether we are overfitting retiring data, then that the error behaviour showed in the graph is clean a induction.

Something particularly interesting is that the use of the very same boast (e.g. in our lotion an indicator, a take profit or stop loss mechanics, etc) might or might non cause overfitting according to the dynamics we are trying to fit.

The reason buns this is that some phenomena (OR some multiplication symmetric variants of the same phenomenon) simply can't be described by any features.

As an example, imagine you are trying to forecast the future sales of adannbsp;sportwear store in Australia. A "good" feature to use could be the season of the year, as (say) Aussies are peculiarly keen in water sports and so springs and summers tend to show the best gross revenue for the year.
Now imagine tryingdannbsp;to forecast the future sales of a similar sportwear store located somewhere in the US. It might equal the case that US citizens don't have a penchant for any particular season, as in the summer they practice water sports and in the winter they go skiing. In that new scenario, a model using the harden of the year as a feature is more likely to result in an overfitted worthy because of the different rudimentary dynamics.

Back to financial markets, an example of this could be how a stop loss mechanism tendsdannbsp;to be (in general speaking and according to my experience) a good feature for trend-following strategies,dannbsp;but not for mean-reversion strategies (and viceversa for objective profit orders). A possible explanation of this could be that trends are well described past the absence of big adverse movements,dannbsp;simply their full extension can't be famed beforehand (but this is just ME nerve-racking to rationalize my empirical findings).

And so,how do you understand which features are good candidates?
Luckily for us, there are a whole cluster of techniques developed in the Machine Learning field to operate feature selection. I advocate the following 2003 paper for an overview of the methods: An Introduction to variable and feature selection asidedannbsp;Isabelle Guyon. Whatsoever school tex of Machine Acquisition should also cover some of the techniques, as it does the exceptional Stanford University's Machine Learning class in Coursera
Some other readers' recommendation (or comment) is course very welcome.

Andrea

Cut public presentation estimators

This is a quick review on my previous post on Quantile normalization.

Or else of removing just the upside X quantile of returns/trades when optimizing a strategy's parameters space, my recent approach has been to remove the top anddannbsp;freighter X quantiles, so effectively using a hardydannbsp;trimmed estimatordannbsp;of performance instead of the figurer itself.

The advantages are symmetric to those discussed in the previous post, as long as your backtest allows for realistic modelling of trades execution – e.g. if dannbsp;you are using stop orders and trade parallel bars (as opposing to tick data), you probably want to summate an amount of slippage in extraordinary way proportional to the size of bar (specification needed because a conservative modelling of limit orders is easier to achieve).

Passementerie out the inferior returns is particular useful just in case of strategies having idiosyncratic big losings (such are signify-relapse strategies of some openhearted usually), whereas trimming the best returns is more useful for strategies with big positive days (e.g. trend-followers strategies).

Two (of many) possible variants are:

-To preserve autocorrelations of a strategy's returns, single could decide to absent blocks of trades/days, rather of individual trades/days (in a similar fashion to what one does when bootstrapping blocks of trades/days).

-To preserve the issue of samples in our results instead of removing the top (worst) days, one could replace them with the average out/median positive (losing) days.

Something else to note is that if your performance measure makes use of std deviation (as IT's the cause for Sharpe Ratio), trimming the tails of the returns from its computation is apt to result in an overestimation of the performance.

Finally, here's the Matlab code:

———————————
normalise_excess_pnl = 1;
normalisation_quantile = 0.98;

if normalise_excess_pnl

best_daily_pnl =dannbsp;quantile(pnl_daily,normalisation_quantile);
worst_daily_pnl =dannbsp;quantile(pnl_daily,1-normalisation_quantile);

pnl_daily(pnl_dailydangt;=best_daily_pnldannbsp;) = [];
pnl_daily(pnl_dailydanlt;=worst_daily_pnldannbsp;) = [];

final stage

———————————

(I usually have got the variable star normalise_excess_pnldannbsp;mechanically initialised to 1 surgery 0 from the external environment, according to whether surgery not I'm lengthways an optimisation).


Andrea