Testing real-time forecasting systems - page 46
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Sergei, why do you build a boundary with one RMS and not 3, for example, for 99.97% probability (if the distribution is normal)? Maybe one sigma is not enough for a reliable prediction?
PS. I understand that the question is rather rhetorical, but hope that the price will fit within one sigma...
I do additional calculations on the margins, but I don't show everything. Conceptually, the model assumes two approaches:
- (1) Estimation of "statistically possible trajectory" (just what I show on the forecasts). That is, it is a kind of "mathematical expectation" (taken in quotes) of the future process as a trajectory. Future realisations will probably form somewhere near from.
- (2) Clustering trajectories and identifying alternative "directions" in the evolution of the System. This is a more detailed analysis, but is still in the works.
An alternative must be understood as a kind of transition from one process model to another, where these models are very different and will lead to significant deviations from the "starting" conditions.It looks like the actual price schedule is rhomno halved, relative to the forecast.
Then the prognosis coincides in time, then in price or even more likely in structure.
Since I am only showing the first point now, this is essentially an "averaging" of plausible process trajectories. Roughly speaking, the RMS is always less, and sometimes substantially less, than the process spread (max() - min() ) And one should not expect an exact hit here. This is only possible in one case (sometimes encountered), - the almost complete absence of alternatives.
So as not to be unsubstantiated, below is a family of the most likely trajectories (not all) for the week from now. EURUSD, watch, forecast 120 hours ahead
this household handles the NS. To the eye, you can see that you have to wait for the downswing, although there is still the possibility of going upwards. Same as always, so there you go, but be careful. :о)
Without any detailed commentary. The "probability vector" is shifting slightly:
So, we are again close to a possible 1.5
grasn, есть маленькое рацпредложение, если оно, конечно, не противоречит вашей личной политике публикации прогнозных материалов: сопровождать каждый рисунок csv-файлом со структурой "время, цена". Это дало бы возможность воспроизвести прогноз непосредственно на чарте в МТ и, может быть, сделать дополнительный анализ по мере развития событий.
no problem. I will report the csv file and maybe have time to do the MT conversion. The only thing I remind you is that this is only testing and you should not completely trust the forecasts. Besides, the situation is changing and on the good side, the forecast should be recalculated more often than I do.
- Amazing, I've created a parallel universe!
Quote to your avatar, nothing personal ))
- Amazing, I have created a parallel universe!
Quote for your avatar, nothing personal )))
Good to see colleague :o) As for the avatar and the quote - it's my favourite cartoon ever. I, for example, also like this one:
-Professor, where were you on the 15th evening?
- (rubbing the back of his head) - And where am I now?
PS: Well, since you're the first to begin, ... just curious, why do you respectable bucket on his head?
Futurama, The Simpsons and Tickle and Scratch, three of the greatest animated series in existence.
(not a lot of wood I've spent collecting all the seasons, especially the Simpsons, they've been around since '86)
^_^
....
I like your forecasts, in my opinion it is quite right that the forecast is constantly "mutating" by digesting fresh data.
Let me ask you a couple of questions:
- Have you managed to automate trading on these forecasts?
- How does the accuracy of forecasts look like depending on TF ("scale" of input data), forecast depth/window? Did you manage to find the golden mean? It's a stumbling block for me...
- And if i may, in two general words and approximately, what is the input of NS?
But my stumbling block is a low-performance computer (Intel(R) Pentium(R) Dual CPU E2220 @ 2.40GHz).
If it were at least 1 000 times faster, I would be able to analyse-calculate H-C-L only at M1, to make a forecast for a couple of days (2880 bars) ahead.
In doing so, the prediction itself should be approximately 5-10% of the analyzed historical data (minimum 30000 х 3 = 90 000 values of H-C-L)
And to optimize the algorithm we need at least 3 months of М1 history (IMHO). Although I missed the point because I am actually making a forecast 8 hours ahead on M15
To calculate the forecast for 8 hours ahead about 10 days of data (about 600-700 prices Close, High or Low separately), to optimize the algorithm - for 1.5 - 2 months of historical data.
Calculation takes approx. 10 sec, optimization - 15-20 min. in deductor. Automation - manually: run one script - calculate the forecast; run another - display the result on the graph.