Voodoo Science-Statistical Analysis | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

[Book Discussion] Thomas R Demark - New Science of Technical Analysis - Chapter 1 30 replies

Best tools for statistical analysis of tick data and news events? 8 replies

Statistical Analysis of Trading History 9 replies

Statistical analysis fallacy 33 replies

The New Science of Technical Analysis book 3 replies

Attachments: Voodoo Science-Statistical Analysis
Exit Attachments

Voodoo Science-Statistical Analysis

  • #1
  • First Post: Jun 5, 2016 10:48am Jun 5, 2016 10:48am
It doesn't matter whether an institutional trader-slash-quant-slash-mathematician thinks she can beat the market for years to come, in backtest. A question remians---does she knows the difference between science, and voodoo science with faulted assumptions?

e.g. Normal Distribution, Expected Value


Any ideas for "Anti-Statistics"?

Got an eye candy for you ↓↓↓
Attached Image (click to enlarge)
Click to Enlarge Name: sigma.png Size: 26 KB
No Highs? No Lows? Must be Bose!
  • #2
  • Jun 5, 2016 11:01am Jun 5, 2016 11:01am
Let me start with a robust measure of things---Median

Attached Image (click to enlarge)
Click to Enlarge Name: median.png Size: 17 KB



Pro: more robust than MA
Con: less sensitive & lack of precision
No Highs? No Lows? Must be Bose!
  • #3
  • Edited 3:04pm Jun 6, 2016 2:44pm | Edited 3:04pm
Quoting Tardigrade
It doesn't matter whether an institutional trader-slash-quant-slash-mathematician thinks she can beat the market for years to come, in backtest. A question remians---does she knows the difference between science, and voodoo science with faulted assumptions? e.g. Normal Distribution, Expected Value Any ideas for "Anti-Statistics"? Got an eye candy for you ↓↓↓ {image}
By plotting a median, are you not working off the faulty assumption that your data is normally distributed? Is it? There are many normal curves, they all share an important property that allows us to treat them in a uniform fashion. The 68-95-99.7% Rule

If your data is not a normal distribution (normally distributed) it is invalid to apply normal/gaussian distribution analysis. (Normal distribution is a sub-form of Gaussian distribution)

Where is the validity of the fit of the distribution used for your analysis (median)? If your data doesn't not follow the assumed (normal) distribution, the results will be inaccurate. Analysis critically depends on the validity of the fit of the distribution used for the analysis

In case of having "non-normal" data the sigma calculation is not effective as the high and low specs are calculated based on Z values.

Your idea supporting , anti-statistics, "excuse me? How many sigma was this?" example comes from your faulty assumption (voodoo science) that your data is normally distributed. Did you check? Or are you working off the faulty assumption that it is?

I'm not sure what kind of point that you are trying make? Other than that performing normal distribution analysis on a non-normal distribution is invalid? That is not "voodoo science", its common knowledge. "Voodoo science" is thinking that performing normal distribution analysis on a non-normal distribution is valid.

Im pretty confident that a "institutional trader-slash-quant-slash-mathematician" knows the difference between science, and voodoo science with faulted assumptions (the difference between normal and non-normal distributions)


The important question is, DO YOU?
Markets are not efficient, rather they are effective - Jones
  • #4
  • Jun 6, 2016 3:13pm Jun 6, 2016 3:13pm
Quoting Tardigrade
It doesn't matter whether an institutional trader-slash-quant-slash-mathematician thinks she can beat the market for years to come, in backtest. A question remians---does she knows the difference between science, and voodoo science with faulted assumptions? e.g. Normal Distribution, Expected Value Any ideas for "Anti-Statistics"? Got an eye candy for you ↓↓↓ {image}
Theoretically you can say about more than 9 sigma.
However markets are not normally distributed.

Best
D.P.
Attached Image (click to enlarge)
Click to Enlarge Name: 2016-06-06_23-10-06.jpg Size: 203 KB
Hunting High and Low
  • #5
  • Jun 6, 2016 10:02pm Jun 6, 2016 10:02pm
Quoting mzvega
{quote} By plotting a median, are you not working off the faulty assumption that your data is normally distributed? Is it? There are many normal curves, they all share an important property that allows us to treat them in a uniform fashion. The 68-95-99.7% Rule If your data is not a normal distribution (normally distributed) it is invalid to apply normal/gaussian distribution analysis. (Normal distribution is a sub-form of Gaussian distribution) Where is the validity of the fit of the distribution used for your analysis (median)? If your data doesn't...
After re-reading my own post, yes my point is clear, but allow me to elaborate further

1. Market returns are not normally distributed.

2. Statistics is a form of data science that collects, analyzes and interprets data. Falsely apply parametric statistics to market return analysis however, is "voodoo science".

3. The ridiculous example of measuring black swan using "sigmas" is exactly the counter-example of how faulted the original assumption of normal distribution could be. Therefore, 68–95–99.7 rule would not work, to say the least. The above example of USDCHF crash was well beyond five-sigmas if using standard deviation assumption, yet a five-sigma shouldn't have happened according to parametric statistics---it would've only happened once in 3.5 million years.

4. A "better" approach to market analysis is to use non-parametric statistics, median happens to be one of the most robust & simplest.
No Highs? No Lows? Must be Bose!
  • #6
  • Jun 6, 2016 10:13pm Jun 6, 2016 10:13pm
Quoting Dr.Phoenix
{quote} Theoretically you can say about more than 9 sigma. However markets are not normally distributed. Best D.P. {image}
Touche
No Highs? No Lows? Must be Bose!
  • #7
  • Jun 7, 2016 2:42am Jun 7, 2016 2:42am
Where is this thread supposed to go? It is known since the late 60's that the markets returns aren't normally distributed. Median assumes the distribution to be symmetrical i.e. no drift (aka trend/momentum). Mandelbrot has fit a Lévy distribution to the markets. This distribution has infinite second moment. How many "sigmas" (of Lévy) is this move? The answer is zero. Thus far I understand: There is no trend and the risk is infinite => markets are worse than casinos.
No greed. No fear. Just maths.
  • #8
  • Jun 7, 2016 2:52am Jun 7, 2016 2:52am
Quoting Tardigrade
It doesn't matter whether an institutional trader-slash-quant-slash-mathematician thinks she can beat the market for years to come, in backtest. A question remians---does she knows the difference between science, and voodoo science with faulted assumptions? e.g. Normal Distribution, Expected Value Any ideas for "Anti-Statistics"? Got an eye candy for you ↓↓↓ {image}
A better question might be on that spike is whether or not that gap in liquidity was going to get filled. There is usually an attempt to fill a price gap. Try not to overthink things if you can help it.
If you are good at something, never do it for free--Joker
  • #9
  • Jun 7, 2016 3:19am Jun 7, 2016 3:19am
Quoting jmn5611
{quote} A better question might be on that spike is whether or not that gap in liquidity was going to get filled. There is usually an attempt to fill a price gap.
Statistically it means the price is recurrent and it implies finite variance. Good news. It also implies a null expected trend. Bad news.
No greed. No fear. Just maths.
  • #10
  • Jun 7, 2016 3:28am Jun 7, 2016 3:28am
Quoting PipMeUp
{quote} Statistically it means the price is recurrent and it implies finite variance. Good news. It also implies a null expected trend. Bad news.
I am definitely in the camp that price is recurrent over a wide enough time frame. Looking at MN charts it looks often like those trends we hunt are actually legs in a larger trading range. I am prepared to hold a position with a time stop of three months based on my very unscientific research into this.
If you are good at something, never do it for free--Joker
  • #11
  • Jun 7, 2016 3:32am Jun 7, 2016 3:32am
Quoting PipMeUp
Where is this thread supposed to go?
Theoretical discussions on whether there are better tools for non-normal data analysis
No Highs? No Lows? Must be Bose!
  • #12
  • Jun 7, 2016 4:32am Jun 7, 2016 4:32am
Conventional trading systems based on parametric assumption are capable of generating countless trading signals based on normal price data, collects paper premium till the day of reckoning

The opposite system is to trim normal data, but generates signals when black swan happens

First you need more robust outliers scout, IQR

Attached Image (click to enlarge)
Click to Enlarge Name: IQR.jpg Size: 376 KB
No Highs? No Lows? Must be Bose!
  • #13
  • Jun 7, 2016 7:52am Jun 7, 2016 7:52am
Quoting Tardigrade
{quote} Theoretical discussions on whether there are better tools for non-normal data analysis
I workaround the problem by studying constant range bar charts to (almost) return to the normality. As soon as you try to solve things with fat tailed distributions you get untractable equations.

Quoting Tardigrade
Conventional trading systems based on parametric assumption are capable of generating countless trading signals based on normal price data, collects paper premium till the day of reckoning The opposite system is to trim normal data, but generates signals when black swan happens First you need more robust outliers scout, IQR {image}
I see a BB on your screenshot. Even if the returns were normal this indicator is fundamentally flawed. It uses the same window as the SMA period (usually 20-50) to estimate the variance while at least 300 samples are required to have an acceptable estimate (<10% error). It doesn't take the lag of the SMA into account.


Something way much more problematic than the fat-tail, to me at least, is the high autocorrelation of the price series. aka pink noise.
No greed. No fear. Just maths.
  • #14
  • Jun 7, 2016 10:33am Jun 7, 2016 10:33am
It is possible to compute the empirical probabilities of reaching a given deviation level without any need to make assumptions about the type of distribution that best represents the market data. For example on EU 15M with a 100 bar moving median calculation to approximate the 68, 95, 99.7 rule with this data the equivalent deviations are approximately 1.5, 3, 4.6 vs the typical 1,2,3 standard deviations expected under normality assumptions. The real question (at least in my mind) is whether you can do anything useful with these numbers once you have computed them?

To show that these probability estimates are fairly robust through time, I split the data series in half and computed the z score probabilities for each half independently. These probabilities for a given Z score can be estimated on half the data, and seem to have some fairly good ability to generalize to the unseen half data with high accuracy: (see pic) the white dots are the first half of the data, and the red dots & lines are the 2nd half of the data. The red and white dots mostly overlap (particularly in the higher deviations) showing that the probabilities of (price - median) staying within a given z score, are pretty consistent over this series. Z score is on the X axis.
Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot2.png Size: 22 KB
  • #15
  • Jun 7, 2016 11:03am Jun 7, 2016 11:03am
Quoting PipMeUp
Median assumes the distribution to be symmetrical i.e. no drift (aka trend/momentum)

Median is supposed to be distribution-free??? Got any papers to back up your claim? Thanks
No Highs? No Lows? Must be Bose!
  • #16
  • Edited 11:37am Jun 7, 2016 11:11am | Edited 11:37am
Quoting PipMeUp
{quote} I see a BB on your screenshot. Even if the returns were normal this indicator is fundamentally flawed.
Haven't used BB for years. The BB I posted above was for demostration only. Nevertheless, a BB with 10 SD is worth a shot.

https://en.wikipedia.org/wiki/Chebyshev's_inequality


[/quote]Something way much more problematic than the fat-tail, to me at least, is the high autocorrelation of the price series. aka pink noise.[/quote]

Why is that?
No Highs? No Lows? Must be Bose!
  • #17
  • Jun 7, 2016 12:06pm Jun 7, 2016 12:06pm
Quoting FXEZ
The real question (at least in my mind) is whether you can do anything useful with these numbers once you have computed them?

/Insert Michael Scott Thank You gif/

It depends on whether you want a robust estimator of location or scale?
No Highs? No Lows? Must be Bose!
  • #18
  • Edited 1:28pm Jun 7, 2016 1:18pm | Edited 1:28pm
Quoting Tardigrade
It doesn't matter whether an institutional trader-slash-quant-slash-mathematician thinks she can beat the market for years to come, in backtest. A question remians---does she knows the difference between science, and voodoo science with faulted assumptions? e.g. Normal Distribution, Expected Value Any ideas for "Anti-Statistics"? Got an eye candy for you ↓↓↓ {image}

Assumption YES but not necessarily faulted.
  • #19
  • Jun 7, 2016 1:21pm Jun 7, 2016 1:21pm
Quoting Tardigrade
A "better" approach to market analysis is to use non-parametric statistics......
I agree.................

This is why I prefer Auction Market Value Analytics. (Not a form of TA)
It works using a frequency distribution; frequency distributions need not be symmetrical. If your data is distributed bounded by 3 price time events using only Market generated info/data. Market generated info/data is the markets transactional data collected and organized so that price and time are displayed.

If you break the data down to its smallest measurable (quantifiable) measure the market consists entirely of price time events (units/tpo's). This is not subjective. Every price reported in the market has a time element. Price and the time in which it occurred. This is market defined.
Attached Image(s) (click to enlarge)
Click to Enlarge Name: amvt1.JPG Size: 246 KB
Click to Enlarge Name: amvtz.JPG Size: 94 KB
Markets are not efficient, rather they are effective - Jones
  • #20
  • Edited 6:58pm Jun 7, 2016 1:44pm | Edited 6:58pm
Quoting Tardigrade
{quote} Theoretical discussions on whether there are better tools for non-normal data analysis
your analysis is only as good as the data being analyzed..............Before you choose a tool, does anyone not care about "data validity"?

Mathematically, the market is recognized as a "non-linear, non-homogeneous differential equation with non-constant coefficients", impossible to solve in closed form. For practical purposes (AMVT) we deal with variables we can observe, price, time and volume. From these come estimated value.

Nowhere outside of AMVT does anyone discuss or address issues of data validity. Is your data fit for the questions being asked? Practice as permitted by theory (AMVT) disallows the use of Bar charts. Linear data. The market is recognized as non-linear. Yet TA uses linear measures in the form of bar chart data (ohlc) for analysis to describe a market that is recognized as non-linear. Measuring one price against another price, linearly.

This is a re-hash of a previous post of mine.
The market is recognized as non-linear, this is obviously illustrated by how price is reported
Attached Image (click to enlarge)
Click to Enlarge Name: tick data.JPG Size: 174 KB


Raw market data is not a linear series. Lets define a bar chart.......

The market is reported as ticks, Price and the time in which it occurred (2D)
What happens next is the data is "subjectively" broken into TF's ( 1m, 5m, 15m, and 30m.etc.) depending on the viewer. This artificially skews the data. Each place where the data has been "subjectively" broken into TF's creates a high and a low. This is "subjective" info because it is biased by the observer's choice of TF. Already you have "flaws" in your data set.

Now that the data has subjectively been broken into TF's. The data inside timeframe is now "sorted" (rearranged) into a linear "series" from high to low forming a bar/candle. Now you have skewed your data set once again. Remember the tick data didnt start as a linear series. Your TF's only look like a series of highs & lows because you manipulated the data to do so. This is not "Market Generated info".

Bar chart data, the high & low was determined by the viewers choice of TF, not set by the market itself, and "price and the time which it actually occurred is removed by sorting/rearranging the data into a high to low series.

A bar/candle only has 4 points of reference. OHLC. Each one of the values is the result of skewed & manipulated data process.

Demand differs at various prices.
A day's recorded summary prices (open, high, low and close/candlesticks) can provide misleading signals: There is typically little trading volume at the highs and lows, so these prices are often not representative of market demand. Likewise, the open and close may well be near the low volume highs and/or lows and themselves may not be representative.

Most traders relying on technical analysis lose, not because of the methods, but because the data used is screwed & chopped before any analysis takes place. Garbage in = Garbage out. Candlesticks, bars prices are often not representative of market demand. This is why TA is often viewed as "Voodoo science".

Demand can be read from the frequency of times a price has been used This is why my analysis comes from frequency distributions rather than a bar chart.

Rule 1 in auction market analysis is "data validity" is your data fit for the task at hand? Bar chart data is skewed and removes the "time" element from price.

AMVT "re-formats" the skewed OHLC data into its original "price time" (tpo) format

EDIT: Auction Market Value Analytics is the quantitative analysis of reference points. It has nothing to do with the holistic pattern recognition used in Market Profile. The only thing Auction Market Value Analytics has in common with MP is that they use the same "price time" format.
It uses the data for monitoring small changes in the market & trade decisions......not the "graphic"
Attached Image (click to enlarge)
Click to Enlarge Name: Auction Market Value Analytics.JPG Size: 338 KB
Markets are not efficient, rather they are effective - Jones
0 traders viewing now