loelschlaeger / fhmm Goto Github PK

Add brackets for functions in flowchart.

• sort: state with highest mu at the front, descending
• design estimation result output in txt-file (names, elements, order)
• Hessian computation
• AIC and BIC computation
• check if iterlim was exceeded, if so, increase

• Set_controls() sollte als zweites stehen, da es für prepare_data() wichtig ist. Klarstellen: An object of class RprobitB_controls oder fHMM_controls?
• prepare_data() sollte als drittes stehen, weil hierbei erst die class fHMM_data eingeführt wird, was man für das jetzt zweite plot braucht. Das könnte verwirren
• Decode_state: ich würde „the most likely” state sequence schreiben. Rprobit_B unklar
• fHMM_events(): mir wäre nicht ganz klar, was diese Funktion genau macht. Prüft sie gegeben Events, dass diese passend sind zum Einlesen?
• fHMM_parameters: A tpm of dimension controls$states[1]. – tpm würde ich ausschreiben. Unterschied mu/mus_star wird nicht klar

Try

controls = list(
  id            = "test", 
  sdds          = c("normal","normal"),
  states        = c(3,2),
  time_horizon  = c(100,30),
  at_true       = TRUE,
  overwrite     = TRUE,
  seed          = 4
)

and see that this gives an odd x-scale in sdds.pdf. Set them based on distribution limits.

Check the new functions set_controls and prepare_data for clear documentation and expected behavior.

First of all - I love the package! I struggled with some not-so-userfriendly packages in the past, but this is really something else!

To my question - is there existing funtion in the package to visualize/plot only the simulated data but with the same structure, visualizing the state scales underneath?

If not - is the simulated data bundled together with the data I fitted the model on in the data.rds file?

In the picture above, logReturns and dataRaw with 2714 elements. I only modelled a year, so this got to be all of the data, right? So I only need to fetch he last 365 elements and then I'm fine?

I do hope I was clear enough. If there are any confusion, just leave a quick comment and I will try to explain further.

Thanks in advance,

Carlos

Incorporate covariates into the state process(es) to determine which factors affect the probabilities of switching to bearish and bullish markets, respectively (just an idea, perhaps something for later versions of the package!).

Create (tikz?) graphic of package functions.

• make visualization more flexible for any number of states
• add. parameter: vector with dates and labels to highlight in the plot (e.g. Lehman bankruptcy)
• check that plots don't get overwritten

• check that nLL_hmm works
• check that nLL_hhmm works
• transformation of thetaUncon to thetaCon with function
• nLL_hmm can also be called from nlm

Use model results to forecast the market.

write funtion "check_controls" that checks control parameters and gives output about model formulation

Problem

Log-return averages on the coarse scale seems not to be the best idea. It's hard for the code to detect different states / state switches for this type of data.

Idea

Include parameter in controls to select type of coarse scale data (e.g. sum of absolute values, mean, average of absolute values). Plot coarse-scale data in ts.pdf to see if this yields better data.

• mini example
• descrp of controls
• saved outputs

Use option hessian=FALSE in nlm and only hessian=TRUE in final estimation run. Should give speed improvement.

Implement coef method to extract estimated model coefficients.

E.g. check_estimation or checkEstimation or check.estimation?

Add #' @export as last Roxygen tag for user-level functions.

• Add a folder called "R" that contains all .R files and a folder called "scr" that contains all .cpp files (this is the folder structure that is required for the R package). Here's a cheat sheet on creating packages that could be useful for the development: https://github.com/rstudio/cheatsheets/raw/master/package-development.pdf.
• Create a separate .R file for each function.
• Write documentation for each function using roxygen tags, see comment below.
• Where functions from other packages are used, use packageName::functionName() instead of functionName() to avoid conflicts.
• Choose name for the package: fHMM
• R package hex sticker
• description file

Give controls parameter "data" which is a list containing all parameters related to data processing/simulation. Update documentation.

Beim Durchlaufen dieses Codes:

simulated HMM -----------------------------------------------------------
seed = 1
controls = list(
states = 3,
sdds = "gamma",
horizon = 500,
fit = list("runs" = 100)
)
controls %<>% set_controls
data = prepare_data(controls, seed = seed)
data %>% summary
data %>% plot
model = fit_model(data, ncluster = 1, seed = seed) %>%
decode_states %>%
compute_residuals
summary(model)
model %<>% reorder_states(state_order = 1:3)
compare(model)
model %>% plot("ll")
model %>% plot("sdds")

wird der 3. Status leider nicht richtig erkannt. Ich habe dasselbe auch mit 1000 Runs einmal ausgeführt, geändert hat sich am Ergebnis allerdings nichts.

Try the fixed-dfs model

controls = list(
  id = "test",
  sdds = c("t(Inf)",NA),
  states = c(2,0),
  time_horizon = c(100,NA),
  seed = 1
)

and see that two states cannot be identified. However, the dfs-flexible model

controls = list(
  id = "test",
  sdds = c("t",NA),
  states = c(2,0),
  time_horizon = c(100,NA),
  seed = 1
)

works.

Hi,

I am trying to fit an hhmm on copper data, however, after

fit_hmm(controls, events)

the following error pops up:

Error in check_controls(controls): File './data/x.csv' not found.

Thank you in advance for the support.

Make sure that the reference style is consistent in ref.bib (to discuss).

• Early stopping of non-promising optimization runs.
• Parallelise numerical optimization runs.
  • Set number of cores in controls via ncores.
  • In check_controls, read out available number of cores, give warning if not (all-1) and error if too many (>=all) cores are used.
  • Divide all runs into ncores batches. Last one has ceiling(runs/ncores) runs, all others floor(runs/ncores) runs. Implement progress bar for last batch. ncores must not exceed runs.

Run github instance, error reported

A collection of ideas on how to further extend the code:

• Functionality to loop over different numbers of states.
• How to deal with NA values in empirical data ("Close" may not exist for every time point, two data sets may not share all close days)?
• Include comparison between true states and predicted states for simulated data in contingency table.
• Possibility to extract any column from dataset, not only "Close".
• Extend for fix of degrees of freedom on one scale only.
• Show progress bar before first iteration
• Give error if any state = 1.
• Download new data automatically from https://finance.yahoo.com/. Write function download_data in 'data.R'.

Write own code for progress + ETA output. Remove dependence to "progress" package.

simulate HMM and HHMM (depending on N=0 or N!=0)

For empirical data, implement that the fine-scale horizon can be monthly / quarterly. Leads to different fine-scale chunk sizes. In this case warning, if !controls[["data_cs_type"]] in c("mean","mean_abs")

Give controls parameter "nlm" which is a list containing all parameters that can be passed to nlm. Update documentation.

Transform graphics from base R to ggplot2.

Gives output %18-s 3 / 2 instead of number of states: 3 / 2

• process two different sources of data
• truncation: find nearest date if truncation point does not exist, possibility to include NA for no truncation
• print out which data is used and how
• check in check_control if correct emp. data is supplied for HMM and HHMM

Try

controls = list(
  id            = "test", 
  sdds          = c("normal","normal"),
  states        = c(3,2),
  time_horizon  = c(100,30),
  at_true       = TRUE,
  overwrite     = TRUE,
  seed          = 4
)

and see that the pseudo-residuals of the fine scale are not normal.

Extend code for other state-dependent distributions:

• t: t-distribution
• t(x): t-distribution with x fixed degrees of freedom (which replaces fix_dfs in controls)
• norm: normal distribution, i.e. t(Inf)
• gamma: gamma-distribution

Include control sdd (character vector of length two).

Set JSS paper structure and metadata.

• Hessian checks (Timo)
• integration into code (Lennart)
• results to output file (Lennart)

I'm doing an analysis of our predictions and I would like to access the output in a list/array so I could run MAPE, MSE and some other indicators. I can't seem to find it - where is it?

• fit Viterbi to new code

add small executable examples in main .Rd-files to illustrate the use of the exported function but also enable automatic testing

I'm raising the issue again - I can't find the simulated data in the provided model files. Just the graphical representation. I've looked through the output files and I can't find it.

See residuals and ts plot.

Document

• thetaUncon
• thetaCon
• thetaList
• *Ordered

in ReadMe. Make sim_par to thetaList object.

Implement function that creates LaTeX output of model formulation and results.

• Add @return to roxygen tags and explain the functions results in the documentation.
• Write about the structure of the output (class) and also what the output means. (If a function does not return a value, document that too).
• Missing @return tag in
  • apply_viterbi.Rd
  • check_decoding.Rd
  • create_visuals.Rd
  • download_data.Rd
  • fit_hmm.Rd
  • plot_ll.Rd
  • plot_sdd.Rd
  • plot_ts.Rd

Include test for normality of pseudo-residuals, see package tseries, function jarque.bera.test.

download_data("dax", "^GDAXI", path=".")

download_data("hk", "HEN3.DE", path=".")

horizon: 2020-01-02 to 2021-03-01

Warnings:

1. possibly unidentified states (C.6)
2. events ignored (V.2)

Ensure that functions do not write by default in the user's home filespace (including the package directory and getwd()). This is not allowed by CRAN policies.