Bitcoin Forum
August 20, 2018, 01:48:49 AM
 News: Latest stable version of Bitcoin Core: 0.16.2  [Torrent].
 Home Help Search Donate Login Register
 Pages: [1]
 Author Topic: Time series analysis  (Read 1223 times)
dree12
Legendary

Offline

Activity: 1246
Merit: 1031

 January 28, 2013, 08:54:11 PM

I am thinking of producing a few short-term forecasts for Bitcoin price, based on chodpada's now-private time series analysis. My goal is to produce a forecast that is more accurate than the null forecast, which I will explain below. To accomplish this, I seek to maintain several "indicators", each with a certain weight that varies based on how well the current situation coincides with a situation in the past.

No experiment is tenable without a benchmark to compare to. The null forecast provides this benchmark. Effectively, the null forecast predicts that the daily average price for any day in the future will be exactly equal to the daily average for the last day in the time series. In other words, the null forecast predicts that the delta in daily average price will be exactly equal to zero. This is a simplified version of both arithmetic and geometric extrapolations of the time series, with linear growth of zero and logarithmic growth to the base of one. If a forecasting method performs better than the null forecast, then it has merit for future consideration. If even the null forecast outperforms a forecasting method, then there are serious concerns with the method and it is likely of little use.

As part of this process, I am employing public opinion. This is to determine the indicators likely to impact changes in price. I have obtained several indicators and their theoretical effects on price, but many of them are likely red herrings.

Because the experts in the Speculation board use indicators all the time, I believe that this public consultation will improve results.

• Money supply. Bitcoin money supply inflation is relatively easy to predict with reasonable accuracy. In theory, with money supply inflation comes decreased price. This indicator has performed terribly in hindcasts—the fastest rates of money supply inflation seem counterintuitively to correlate with price increases, not decreases.
• Time. Bitcoin seems to gain value over time, when averaged in the long-term. This indicator, if modelled properly, will likely be one of the heaviest-weighted.
• Momentum (a.k.a. trend). In Bitcoin's history, momentum, especially upwards, has been a double-edged sword. Although an high upwards momentum usually implies price gains, on certain occasions they have implied unprecedented price losses.
• Moving Averages. When possible, commodities tend to return to the moving average as part of a corrections cycle.
• High/Low. Even when forecasting on a daily scale, intraday events are not to be ignored. The highs/lows of the recent time series could imply reversals or continuations in trend.
• Volume. Higher volumes seem to counterintuitively imply more change.
• Delta volume. If volume is decreasing, it may seem that a reversal is underway. If volume increases, the trend may continue for a while yet.

Any thoughts on these example indicators? Any other indicators that could possibly be useful?
1534729729
Hero Member

Offline

Posts: 1534729729

Ignore
 1534729729

1534729729
 Report to moderator
1534729729
Hero Member

Offline

Posts: 1534729729

Ignore
 1534729729

1534729729
 Report to moderator
Advertised sites are not endorsed by the Bitcoin Forum. They may be unsafe, untrustworthy, or illegal in your jurisdiction. Advertise here.
1534729729
Hero Member

Offline

Posts: 1534729729

Ignore
 1534729729

1534729729
 Report to moderator
bb113
Hero Member

Offline

Activity: 728
Merit: 500

 January 29, 2013, 02:37:53 AM

Your null forecast reminded me, there was someone awhile back who randomly bought or sold based on a coinflip for a couple months. I'm not sure what happened in the end but that thread may be of interest to you. Also, has the daily average ever stayed exactly the same for two consecutive days? That seems like a null hypothesis that is known to be false a priori, but maybe it still works for these purposes.  These are good threads btw.
notme
Legendary

Offline

Activity: 1904
Merit: 1001

 January 29, 2013, 04:59:48 AM

Your null forecast reminded me, there was someone awhile back who randomly bought or sold based on a coinflip for a couple months. I'm not sure what happened in the end but that thread may be of interest to you. Also, has the daily average ever stayed exactly the same for two consecutive days? That seems like a null hypothesis that is known to be false a priori, but maybe it still works for these purposes.  These are good threads btw.

He was flipping for trades on Bitcoinica.  Zhoutong ate his account along with everyone else's (not trading loses).

https://www.bitcoin.org/bitcoin.pdf
While no idea is perfect, some ideas are useful.
12jh3odyAAaR2XedPKZNCR4X4sebuotQzN
 Pages: [1]