Guys, this is definitely not rocket science! My article above is directly based on

https://www.quantstart.com/articles/hidden-markov-models-for-regime-detection-using-r

This reference is also given in the article above (References)! Maybe your read this article first.

Hence, I did not invent this approach! We had the discussion in the german subforum last year, whether it is possible to do technical price analysis in a more profound way. My suggestion was, to use a HMM. And someone had to do the work. I only applied the methodology from the mentioned article to bitcoin price data. And I described, explained, discussed and extended the methodology a little bit further (also for myself).

This is also the reason, why the results I present here, are preliminary and uncertain. At

https://github.com/trantute/sentiment-hmm

you find corresponding R-scripts, with which you can play around. The idea is now, whether other people can reproduce my results or correct it.

More fun with HMMs (https://bitcointalk.org/index.php?topic=5142331.msg51069013#msg51069013):

Ich habe noch schnell die gleiche Analyse wie heute morgen für Dienstag nächste Woche durchgeorgelt und Wahrscheinlichkeiten für bestimmte Preise berechnet:

Der Mittelwert der Simulationen liegt bei ca. $ 9200 Dienstag nächste Woche.

Die Wahrscheinlichkeiten für einen Preis unter ...

• ... $ 9000 ist 35.4%;
• ... $ 8750 ist 20.8%;
• ... $ 8500 ist 10.3%;
• ... $ 8250 ist 4.2%;
• ... $ 8000 ist 1.4%.

Die Wahrscheinlichkeiten für einen Preis über ...

• ... $ 9250 ist 48.0%;
• ... $ 9500 ist 32.1%;
• ... $ 9750 ist 19.3%;
• ... $ 10000 ist 10.4%;
• ... $ 10250 ist 5.0%.

Clearly, one can do a similar analysis directly for the next week so we can put the whole analysis to the (weak) test. The following simulations were started at a price of $ 8000:



The log10-distribution of the sampled prices looks gaussian, with mean =  3.96475 (= $ 9220) and standard deviation = 0.02797357. See the corresponding distribution below (keep in mind that the x-axis is shifted to the right and the scales are not fully comparable due to manually screenshot):



As a result, the probability, that next Tuesday the price lies below ...

• ... $ 9000, is 35.4%;
• ... $ 8750, is 20.8%;
• ... $ 8500, is 10.3%;
• ... $ 8250, is 4.2%;
• ... $ 8000, is 1.4%.

And the probability, that next Tuesday the the price lies above

• ... $ 9250, is 48.0%;
• ... $ 9500, is 32.1%;
• ... $ 9750, is 19.3%;
• ... $ 10000, is 10.4%;
• ... $ 10250, is 5.0%.

Provided, there is no major flaw in the statistics!

Einspruch euer Ehren! (Wobei ich $ 760.000 auch für etwas übertrieben halte, aber mal schaun, vllt finde ich noch einen Bug)

Mit der Methode läßt sich zumindest (etwas) genauer als mit (einfachen) Randomwalks quantifizieren, wie wahrscheinlich es ist, dass McAffee seinen Pillemann behalten kann. (Es ist aber trotzdem analog zu Randomwalks.)

Und aktuell liegt diese bei etwas unter 50% (Pi mal Daumen, da ich die Zahlen gerade nicht zu Hand habe).

Steht auf meiner Todo-List ! Desweiteren werde ich jetzt zusätzlich zu den HMM-Plots immer dieses Histogramm für die erste Woche von 2020 mit angeben. Einfach um zu gucken, wie sich die Preisvorhersage über die Zeit ändert. Ich hoffe, dass ich nicht allzu gravierende Bugs im Code habe.

Man muss bei der ganzen Geschichte auch immer bedenken, dass es bei solch einem HMM auch immer eine "phase transition" geben kann. D.h. wenn z.B. ein Wal plötzlich, während eines Bullenmarkts, auf die Idee kommt seine Coins zu Geld zu machen (mit krasser roter Kerze), dann kann die gesamte Statistik über den Haufen geworfen werden. Das ist letztes Jahr mindestens einmal passiert. Z.B. war der Preissturz letztes Jahr auf $3000 so im damaligen Modell nicht vorgesehen und deshalb sehen die Plots seitdem anders aus. Jetzt ist das logischerweise drin.

Der Preis wird ja immer noch von Menschen gemacht und deshalb ist so ein Modell immer nur eine grobe Annäherung an die Realität bzw. die Realität kann das Modell jederzeit zunichte machen.

Ich habe ein wenig mit dem HMM gespielt und versucht den Preis für die erste Woche 2020 zu simulieren. Das Resultat seht ihr hier:



D.h. im Median haben wir nächstes Jahr einen Preis von ca. $ 760000.  

Hier steht noch etwas mehr dazu:

https://bitcointalk.org/index.php?topic=5142331.0

I used the parameters, which were computed by the 4-state-HMM yesterday (transition matrix and distributions of states, I will explain more about this approach in the next days in the article above, for now, check simulate.R at https://github.com/trantute/sentiment-hmm, if you are interested), and simulated 10^6 corresponding weekly Markov Chains. Each Markov Chain was ...

• ... stopped after 33 steps (33 weeks), which corresponds to the beginning of 2020;
• ... started in the same state (the current state, which is the green node in the graph in the article above);
• ... started at a price of $ 7000;

The return can be transformed to a price again, if a price offset (at the beginning: $ 7000) is given. As a result, the 10^6 Markov Chains "randomly" return 10^6 Bitcoin prices for the first week of January of 2020. Since a Markov Chain is a kind of sophisticated die, the result is nothing else than a price distribution, which is summarized in the histogram below (log10 of price is given on the x-axis, the frequency of prices in a price interval is given on the y-axis). Note, that the median Bitcoin price in 33 weeks is roughly

$ 760.000,  

that is, in 50% of the simulated chains, the price lies below, and in 50% of the chains, the price lies above this value. Further note, that in 0.7% of the chains, Bitcoin dies out (price becomes negative).



DISCLAIMER: THESE RESULTS ARE QUITE NEW AND MIGHT STILL BE BUGGY!

2013 waren es 10er Schritte;
2017 waren es 100er Schritte;
2019 sind es 1000er Schritte;

Es ist sicherlich richtig, dass solche Steigerungen nicht ewig so weiter gehen können und werden, aber bei 21 Millionen Einheiten un 8 Milliarden Menschen ist noch Luft nach oben.

Ich gehe bei meiner Milchmädchenrechnung immer von einem neuen ATH aus. Bei Deiner Rechnung gäbe es kein neues ATH, es gäbe nur ein Zwischenhoch bei ca. 12000$ mit darauf folgender Korrektur.

Ich sage ja nicht dass die $400k schon bis Jahresende erreicht sind.

Immer dran denken Leute: die 7000-8000$ von heute entsprechen ungefähr den 450$ vor drei bis vier Jahren. Das Preisniveau gab es ungefähr im Zeitraum des Jahreswechsels von 2015 auf 2016.

FOMO ist noch lange nicht angesagt. Erst bei 100000$ geht das vielleicht los, was in etwa 6000$ in 2017 entspräche (5-fache des alten ATHs). Und dann kann sich das nochmal vervierfachen!

Im Extremfall ja, aber man könnte sich immer auf die letzten 14 Tage u.ä. beschränken. Eine Börse, die den Hack erst nach 14 Tagen erkennt ist wohl so oder so zum scheitern verurteilt.

Desweiteren gilt, ob nun bei PoW oder PoS, und unabhängig davon, dass imho schon ein mehrstündiger Rollback technisch aussichtslos und reine Stromverschwendung ist, dass sich die Miner/Minter ins eigene Fleisch schneiden, wenn sie die Glaubwürdigkeit der Blockchain unterminieren. Es macht keinen Sinn, die Blockchain durch Konkurrenz abzusichern (minieren), um dann aber über die Hintertür doch noch ein Kartell zu bilden (unterminieren). Die zentralen und somit unsicheren Börsen selbst sind nur Hilfskonstrukte, welche mangels Alternativen auf das Bitcoin-Ökosystem draufgesetzt wurden. Bitcoin kann, mit der richtigen dezentralen Technologie, auf solche Börsen verzichten und trotzdem den Wertfluss von Fiat zu Bitcoin und vice versa abbilden. Jeder Miner mit Verstand weiss dies und weiss auch, dass es unfähige zentrale Börsen nicht wert sind durch einen Rollback gerettet zu werden. Natürlich kann ein Miner auch vom Hack der Börse betroffen sein, aber eben nicht alle Miner.

Bitcoin braucht eine glaubwürdige und sichere Blockchain. Börsen brauchen Bitcoin. Aber Bitcoin braucht keine zentralen Börsen.

Oder noch kürzer:

Die Miner schulden den Börsen gar nichts, die Börsen den Minern und Fullnodebetreibern aber alles.

Roughly speaking and given chart data, the idea is, to assign states (= "sentiments") to each week (shown as colored band at the x-axis). For instance, in the plots above, I assume four states, i.e. four sentiments (bearish, sideways, bullish and bubble), which govern/generate/explain the chart. Firstly, from price data of a chart, one computes the return per week. The assumption is, that the return can be generated by four different gaussian distributions, whose parameters are not known, yet. Secondly, an HMM is fitted to the return data revealing the parameters and states, which are most likely.

As a result, one knows quantitatively, at which time point, which sentiment took place. This even works for the current week, even though the current state might not be stable (see robustness analysis). In the moment, the sentiment is bullish.

The idea is explained in the text. I recommend to read it from the beginning. For good reason, I use a simple board game to illustrate the method (https://en.wikipedia.org/wiki/Mensch_%C3%A4rgere_Dich_nicht).

Wie das bzw. woher weisst Du dass dies seine waren? Oder war das Ironie?

Hmm ... es stellt sich die Frage was man möchte. TA, so wie sie in den Foren vorgelebt wird, nimmt für sich in Anspruch, die Zukunft vorhersagen zu können.

Imho geht sowas nicht. Eine mathematisch fundierte Methode kann jedoch Wahrscheinlichkeiten den einzelnen Szenarios zuordnen. Wie und ob man mit diesem Wissen handelt ist dann jedem selbst überlassen.

BTW, ich halte auch das berühmte Bauchgefühl für eine statistische Methode unabhänig davon, dass sie schlecht zu formalisieren ist.

Da möchte ich doch auf meinen Thread zu Hidden Markov Models verweisen (jetzt auch in englisch):

https://bitcointalk.org/index.php?topic=5142331.0

Ein HMM ist zwar auch nur Zufall, aber Zufall in verschiedenen Zuständen ...

Nunja, der gute Zidane. Der hat ein Ego wie Planet. Das Problem mit ihm war, dass er auf der Spitze des Hype 2017 erst behauptete, dass Bitcoin noch die 100k erreichen würde, um dann, als es begann abwärts zu gehen, zu sagen, dass Bitcoin technisch überholt wäre und man doch möglichst alles verkaufen sollte. Die Aussagen ware zeitnah, aber nie VOR den entsprechenden Ereignissen. Und einige seiner Aussagen wurden von anderen Forenmitgliedern mindestens als Halbwahrheiten entlarvt.

Das Problem im Internet ist, dass jeder alles behaupten kann ohne Beweise bringen zu müssen (siehe Craig Wright). Weisst Du denn wirklich, dass er all-out und all-in gegangen ist? Hat er seine Transaktionen öffentlich gemacht? Und so entstehen Urban Legends, welche so gar nicht stattgefunden haben. Es gibt Leute hier im Thread, die haben relativ früh den Boden beziffern können: 3k. Wenn ich mich richtig entsinne waren das qwk und bm42.

Medienkompetenz ist die Fähigkeit nicht alles zu glauben bzw. alles zu hinterfragen, das einem weißgemacht wird. Immer und überall!

The HMM for today - still bullish:

The HMM from last week:

The following text is a translation of one of my comments: https://bitcointalk.org/index.php?topic=4586924.msg41380372#msg41380372. The translation was done quickly but will improve with time.

Prerequisites

• Installation of R
• Installation of necessary R-packages

Introduction

• Example with ideal and manipulated die
• Discussion of resulting variables
• Comparison and selection of models
• Limit of a model

Hidden Markov Models for sentiment analysis

• A price-return Hidden Markov Model
• Robustness analysis
• Other variants

Trading by means of a Hidden Markov Model
Price forecasting with a Markov Chain
Discussion
Outlook
References

Prerequisites

Installation of R

R is a free programming language, which specializes on statistical computing (https://en.wikipedia.org/wiki/R_(programming_language)):

https://www.r-project.org/

This is the only software package needed so that the following code snippets should work by copy and paste. Additionally, it makes sense to install Rstudio:

https://www.rstudio.com/

Rstudio is a graphical user interface and programming environment for R and simplifies programming in R as well as the managment of data.

Installation of necessary R-packages

The following code, executed in R, installs packages, which are necessary or might be useful at later timepoints.


Packages are loaded in R as follows:


Introduction

Example with ideal and manipulated die

The following example is chosen in such a way, that it is analogous to chart analysis.

Suppose you take part in the Mensch Ärgere Dich Nicht World Championship (DNGAWC in the following, which stands for the english translation: Do Not Get Angry World Championchip). For this game you need a die. If you throw a six, you may throw the die again. Furthermore, the player with the six has the choice to come out or throw again normally. Thus, throwing a six is ​​an advantage. For an ideal die, the probability is exactly 1/6, as well as for the other numbers. You could try to help luck via using a manipulated die, which looks as similar as possible to the official die of the world championship. Hence, it would be good, if the organizers of the DNGAWC would be able to detect whether an official die was used or not. Cheating players would then be disqualified from the championship.

At the beginning we consider the ideal die and compare it with the manipulated die. The manipulated die should increase the probability of throwing a six. We assume in this example, that the probability for a six with the manipulated die is 2/7 and for the other numbers 1/7 each. We now throw both dice very often and note down the sequence of the thrown numbers:


To get a feeling for the distribution of the numbers, we combine 500 throws via a moving window on both sequences respectively (also compare other window sizes). We note the mean value for each window and plot the corresponding mean value distribution for both sequences:


The result is shown in the plot below:


Obviously, the two distributions are gaussian. If a sequence was created with the ideal die, then the mean of 500 throws should come from the left distribution. If a sequence was generated with the manipulated die, then the mean comes from the right distribution. If a sequence was created via both dice, then the mean should come either from one or the other distribution. Whether a sequence was created with one or two different dice, i.e., whether the dice were exchanged during the game, can now be detected with a Hidden Markov Model (HMM). However, an important requirement for this example is, that the number of consecutive throws with each die is large (e.g., 10000) compared to the window size (500).

The HMM modelling one die consists of one node and an one edge. In the case of the ideal die, the node corresponds to the upper blue normal distribution. The node generates numbers such that the distribution of the mean from 500 die throws follows the blue normal distribution. Since there is only one edge with the probability P = 1.0, the new state of the HMM after each roll MUST be the old one:


If the ideal die is exchanged by a manipulated die during the game, the corresponding HMM consists of at least two nodes. A node that generates the distribution of the ideal die and a node that generates the distribution of the manipulated die:


The edges with the probabilities P(blue to blue), P(blue to red), P(red to red) and P(red to blue) depend on how often the dice were exchanged during the game. Here, the sequence of numbers should be long enough that more than two exchanges (state changes) took place! The more the better.

An HMM whose parameters are known is also called Markov chain. A Markov chain, interpreted as a model for a given (stochastic) process, mimics that process mathematically and thus represents something like a simulation of the process. In our example, this would be the simulation of the numbers thrown by a fair player or the numbers thrown by a cheater during the DNGAWC.

Thus, an HMM is an automaton, i.e., a graph with several nodes (states) and edges (state changes). One of the nodes is special since it determines the state the automaton currently is in. The edges are labeled with probabilities so that the sum of the outgoing edges sums up to the probability of P = 1.0. The automaton then jumps in discrete steps from node to node along the edges, according to the probabilities indicated at the edges. Since the sum of the probabilities at the outgoing edges is equal to 1.0, the automaton changes its state in each step, though the new state can be the same as the old one. If the HMM is in state A and with the next jump, the state changes to B (with A! = B), then you have a (real) state change.

In the case of an HMM, the parameters of the model, i.e. the probabilities at the edges of the automaton, are not known in the beginning and must be deduced from given data. This is also called fitting the model to the data and the result is simply called the fitted HMM. Having found the optimal parameters, one can now compute the most probable path over the nodes over time through the graph. The sequence of nodes that the automaton passes through is called hidden states (hidden/unknown states). In our example, the data ONLY consist of the thrown die numbers. The hidden states correspond to the use of the dice, i.e. when an ideal die and when a manipulated die was used in the sequence of thrown numbers.

We now generate a simulated sequence of thrown die numbers, which satisfies our requirements (many thrown die numbers in a row by the same die and many dice exchanges), and try to detect the dice exchanges in the resulting number sequence by means of a fitted HMM:


The generation of the HMM, the fit of the model to the data and the extraction of relevant data is a three liner in R:



For the generation of a meaningful plot, we need more code:


The result should look as follows:


On the x-axis, the sequence of states is shown as colored horizontal band. The curves in the plot represent the (a posteriori) probability for the corresponding die at the given time point. Note that the HMM does not reproduce the states perfectly (see also other window sizes). This is due to statistical fluctuations, which means that the blue distribution may look like the red by coincidence (and vice versa). Because of the window, of course, there are less than 180,000 states (179501). Due to numerical inaccuracies, the real transition probabilities can only be accessed using the command "hmmfit@transition". The command "summary(hmmfit)" is not sufficient, because it apparently rounds the probabilities! The corresponding HMM is then given by the following graph:


Discussion of resulting variables

Todo: Discussion of resulting variables

Comparison and selection of models

If the organizers suspect a player to cheat during the DNGAWC, then they must also be able to quantify their suspicion. They must be able to quantify how likely it is that the suspect uses a manipulated rather than an ideal die. In the simplest case, one compares two different models of the suspect: The null hypothesis (H0) is, that the suspect uses an ideal cube; The alternative hypothesis (H1) is, that the suspect uses an additional manipulated die. H0 is discarded if H1 is much more likely. Of course, one can still generate additional models of increasing complexity (H2, ..., Hn), e.g. to model the assumption that the suspect has used multiple different manipulated dice with different parameters. One of those models must be selected in such a way that the probability is minimized to exclude an innocent player from the world championship.

Just as graphs can be varied in their parameters, i.e. in the number of nodes, edges, and thus in their topology (structure), different HMMs can be generated (H0, ..., Hn) and can be fitted to the same data, assuming, that they model reality, respectively. Each fitted HMM will have a certain likelihood. The likelihood here is the "probability" of the HMM given the data and it can be interpreted as a kind of probability. There are also other concepts that are analogous to the likelihood. The best known are the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Given an unfitted HMM, there is virtually always only one fit of the (hidden) states (or a behavior of the HMM automaton over time) to the data, which is the most likely. However, the other fitted HMMs might be similarly likely, even though the absolut value of their likelihood is worse compared to the best fit. In such a case you have to decide which model to choose for further analysis to prevent overfitting.

The naive approach consists of simply choosing the HMM that has the highest likelihood. However, this approach has the problem that a model with more parameters usually ALWAYS has the higher likelihood. At the same time, however, the model loses its general meaning (overfitting) or becomes more difficult to interpret. This means for the set of HMMs to be examined, that the corresponding model parameters, i.e. the number of states, are chosen to be as "reasonable" as possible. However, even in this scenario, the choice of the HMM with highest likelihood can be poor, since simpler models with fewer parameters, that is, HMMs with fewer states, may have "similar" high likelihoods. In this situation, one should choose the HMM with fewest states from the set of HMMs with "similar" high likelihood (Ockham's Razor). How to do that in detail, you can find here: https://stats.stackexchange.com/questions/81427/aic-guidelines-in-model-selection.

Even with new data, the likelihood of one and the same HMM can change, creating a completely different fit. The Expectation-Maximization algorithm (EM algorithm) guarantees to find local maxima but it does not necessarily find the global maximum (sequence of the calculation of an HMM fit: EM algorithm and then Viterbi algorithm). If one fits a model with different random seeds, one might end in different local maxima. In this situation, the random seed can be interpreted as a new parameter and this case can be treated as described in the last paragraph. (Source: https://bitcointalk.org/index.php?topic=26136.msg45063876#msg45063876)

Limits of a model

I do not want to talk much about the limitations of a model. However, I link a documentation where imho everything important is said (The Midas Formula). Although the video does not explicitly refer to the limits of models, but the basic message should be clear in the end. It is important to fully watch the video:

https://vimeo.com/28554862

Hidden Markov Models for sentiment analysis

A price-return Hidden Markov Model

The return of a specific time interval t, e.g. of a specific week, is the difference between closing price and opening price divided by the closing price:


If the closing price is greater than the opening price, the return is positive, if the opening price is greater than the closing price, the return is negative, and if the closing price is equal to the opening price, the return is zero. The return of an asset that continuously varies by a certain average (e.g., like tethers around US $ 1), with no medium to long term swings down or up, can be modeled with a (normal) distribution of mean zero. The return of an asset where, on the other hand, there are additional market situations in which the price also falls or rises in the long term (e.g. bear and bull market in Bitcoin), it makes sense to model these three market situations with at least three different distributions: the distribution of a bear market has a negative mean, the distribution of a bull market has a positive mean and the distribution of a sideways movement has a mean of zero. Those three distributions are equivalent to three different dice at the DNGAWC: a manipulated die where the probability of throwing a six is ​​less than 1/6, an ideal die and a manipulated die where the probability of throwing a six is greater than 1/6.

At https://github.com/trantute/sentiment-hmm, two exemplary HMMs are implemented. The first HMM models four states: bearish, sideways, bullish and bubble (see the plot below). The second HMM models five states: dead, bearish, sideways, bullish and bubble.


The states of the HMM over time are shown as colored horizontal band on the x-axis in the plot above. The posterior probabilities as well as the logarithmized price (normalized to 1.0) and the logarithmized MA203 (normalized to 1.0) are shown as curves in the plot. The sentiment is indicated by color, as described above as well as defined in the legend on the left side in the plot. The corresponding graph is given below:


Robustness analysis

The question arises how trustworthy the sentiment prediction for the current week actually is. Simulations with historical data have shown that the predictions are relatively stable if the market is not volatile in the medium to long term (sideways phase) or in a medium to long-term downward or upward trend (not shown). However, simulations of historical data as well as real data have also shown that an extreme change in price, which has a significant impact on the weekly return, can subsequently change the sentiment prediction. Below, the fits of seven consecutive weeks are shown:


The states of the HMM over time are shown as colored horizontal band on the x-axis in the plot above. As can be seen on the right side in the plot above, a state is unstable, in particular, if the HMM has just changed its state recently, or if different states for the current week have a similar likelihood. The posterior probabilities for the different states (the curves in the plot) can provide information about this. If all states have posterior probabilities not equal to 0.0 - as in the upper plot on the right - then the current state is more unstable than if one state has the posterior probability close to 1.0 and the others close to 0.0.

Other variants

The return as proxy for the sentiment is only one possibility for an HMM. Other variables can be used too, insofar they show a correlation with the sentiment. E.g. the comments per time interval at bitcointalk.org could be used as proxy. Corresponding code can also be found at https://github.com/trantute/sentiment-hmm.

Trading by means of a Hidden Markov Model

If the sentiment is known or if one is able to estimate the sentiment sufficiently accurate, then this knowledge can be used to act accordingly. The idea is to go long at the beginning of an uptrend and short at the beginning of a downtrend. For the plots shown above, this means for example: bearish (red) => short; sideways (black) => long; bullish (green) => long; and bubble (violett) => long. In addition, the plot contains the logarithmized Bitcoin price (US $, Bitstamp, normalized to 1.0) for the corresponding time points (blue) and the MA203 (Moving average for 203 days = 29 weeks, orange) which corresponds approximately to the MA200 but is easier to calculate.

Of course, trading only makes sense, if the sentiment changes from black, green or purple to red or vice versa. For this you have to keep an eye on two weeks each, i.e. suppose today is a Monday, yesterday was Sunday and the HMM has predicted a change of state relative to the previous Sunday. Now you exchange your asset according to the rules mentioned. Namely to the opening price of the current week, so the current Monday. (Source: https://bitcointalk.org/index.php?topic=26136.msg43234811#msg43234811)

Todo: Code

Todo: Test with historic data

Exchange conditions can be additionally optimized. The example from above is quite naive: 0:100, 100:0, 100:0 and 100:0. This are just four parameters. One could search the space of exchange ratios using historic data for finding the ratios maximizing the yield. E.g. steps of 0.1 result in only 10000 possibilities and one choses the ratios which maximizes the yield (in dollar or BTC). In particular, assuming that one can ONLY go long or short, i.e. all in or all out, the computational effort is manageable, since the exchange ratios would then be binary variables. With n = 4 states there would be only 2^4 = 16 possibilities. Such a method may also be useful in finding the best parameter set of the HMM, but with the risk that the model loses generality by overfitting on historical data. (Source: https://bitcointalk.org/index.php?topic=26136.msg43256510#msg43256510)

Price forecasting with a Markov Chain

Todo: Explanation, how to do this. See also https://github.com/trantute/sentiment-hmm for a 4-state-HMM. Also simulations with historic data are needed.

Discussion

It is important to note, that the characteristics of the underlying probability distributions might continuously change over time (phase transition)! For instance, the return probability distribution of a bearish sentiment might be different in 2014 and 2018. As a result, over long time periods, one has to experiment, whether additional states, which model the same sentiment at different points in time but with different probability distributions, improve the power of the model.

Outlook

Todo: Ideas come in here. E.g., price/return forecast via a fitted HMM/Markov chain.

References

https://www.wikipedia.org/
https://api.bitcoincharts.com/v1/csv/
http://web.stanford.edu/class/stats366/exs/HMM1.html
https://www.quantstart.com/articles/hidden-markov-models-for-regime-detection-using-r
https://www.fool.com/knowledge-center/how-to-calculate-return-on-indices-in-a-stock-mark.aspx
https://stats.stackexchange.com/questions/81427/aic-guidelines-in-model-selection
https://github.com/trantute/sentiment-hmm

HMM für heute:



Die wichtigen Informationen sind dem bunten Streifen auf der x-Achse zu entnehmen. Wer will kann die Farben (= Sentiment) mit der Preiskurve abgleichen. Zu beachten ist, dass wir uns weiterhin in der bullischen Phase befinden. Wichtig aber auch, dass jeder extreme Drop mit Einfluss auf den Wochenpreis den HMM nachträglich ändern kann. Trotzdem ein gutes Zeichen, dass die Ampel mal wieder auf grün schaltet.

Letzter geposteter HMM: https://bitcointalk.org/index.php?topic=26136.msg50905954#msg50905954
Quelle und Code: https://bitcointalk.org/index.php?topic=4586924.0