Me, CATS and Spot Gold ... first try

For discussion of how to use the Cycles Research Institute' own Cycles Analysis & Time Series software.

Re: Me, CATS and Spot Gold ... first try

Postby RayTomes » 21 Nov 2012, 01:50

It doesn't look good to me. It is true that sometimes in a long data fit there are some zones that go the opposite way. But for a short series like this I would not entertain that thought.

Next, I would comment that 20 years data is not enough to find some of the main cycles with any degree of certainty. The CATS system comes with several very long gold series, but they are both annual series. So you could look for cycles in the 3-12 year range (say) with those series and then use the results in this analysis. The long term series show a number of cycles including 5.52 years which is close to a figure that appears in many commodities also.

Here is what I would do with your gold data for analysis. Firstly, I saved it as 72-gold.cats in my monthly data directory.

1. When I try
F72 G
I see that the price ranges from about 250 to 1800 which is a wide percentage range. In such cases I normally take logs of prices so that equal percentages become equal changes. (For commodities, the lambda function is another alternative to logs. Recommended reading http://cyclesresearchinstitute.wordpres ... -function/ ). So
F72 # G
and you can see that now the short term fluctuations are about equally much at the low and high prices.

2. There are two ways to go here. Do a spectrum on the log gold prices, or on the rate of change of these. If you decide to use rate of change then remember the way to calculate that from the original
L :1 -
which gives monthly changes (of the log) in this case. You can reverse this later by doing
A
although you may need to add a constant to get the original series values back.

3. For log gold price I get many significant cycles, but they are mostly very short. I feel rather unhappy doing an analysis using data that has such a strong uptrend in it as this does. Warning bells.

Code: Select all
no.cycles  period   date/time maximum  amplitude  Bartels p
87.83   2.85779   727602.109466106   0.006529   0.00018064
49.82   5.03814   727646.006090318   0.019568   0.00023743
70.82   3.5442   727617.662492542   0.015101   0.00044025
35.95   6.98192   727707.071347667   0.02532   0.00131043
38.99   6.43755   727707.666878105   0.024103   0.00152122
111.93   2.24247   727605.124183078   0.006687   0.00154547
34.82   7.2085   727684.134612755   0.020151   0.00200973
57.96   4.33057   727650.380770803   0.014892   0.00276281
30.16   8.32228   727791.816544353   0.030251   0.00292596
19.86   12.63847   727783.078133725   0.037065   0.00421828
94.93   2.64405   727624.399445529   0.005844   0.00452556
110.93   2.26269   727607.163419725   0.009066   0.00470789
90.98   2.75885   727620.753216718   0.009689   0.00482287
33.96   7.39105   727728.996707377   0.026614   0.00560853
68.95   3.64032   727637.348887243   0.004923   0.00642129
101.91   2.46296   727608.952544147   0.015745   0.00652248
22.95   10.93682   727796.843551304   0.041678   0.00679759
24.97   10.05206   727768.739918535   0.022766   0.00704906
73.88   3.3974   727626.999974579   0.007226   0.00729802
66.99   3.74683   727644.100294625   0.010238   0.0080734
58.97   4.2564   727650.390566893   0.004655   0.00953515


4. Try also a spectrum using the log price gold differences
Z F72 # L :1 -
then L S so as to keep the data we are trying to fit on the stack too

Code: Select all
no.cycles  period   date/time maximum  amplitude  Bartels p
70.55   3.54359   727704.172383199   0.182872   0.0139763
46.82   5.3396   727745.477598276   0.169053   0.01435792
101.57   2.46136   727678.407972768   0.22219   0.01544232
87.78   2.84803   727681.717026972   0.200592   0.02312448
86.36   2.89486   727662.899870306   0.18098   0.02551336
45.53   5.49089   727728.186945825   0.161287   0.02657659
84.07   2.97371   727655.246693474   0.178661   0.02920969
104.52   2.39189   727614.66657393   0.19493   0.03045038
98.07   2.5492   727669.511700302   0.187862   0.03310142
93.5   2.6738   727636.063988865   0.15883   0.03794986
92.41   2.70533   727642.856979203   0.187257   0.04834839


There are less significant cycles. There is one of 64.1 months = 5.34 years at p=.0896 so not significant. But is near to the 5.52 year cycle found in longer series.

5. Using only cycles found in both the spectra as a precaution, but including the the 64.1 month one also I did a regression and prediction for the rate of change series. Then I used the "A" command to accumulate both to get back to log price. There is a slight resemblance, but I would not recommend it to anyone.

My conclusion is that there is insufficient data here to make useful predictions.

Might I suggest that we look at your other data (Is it Euro/Dollar?) with the much longer data series? I think that something useful might be done there.
User avatar
RayTomes
 
Posts: 128
Joined: 02 Aug 2010, 23:24
Location: Auckland, New Zealand

Re: Me, CATS and Spot Gold ... first try

Postby RayTomes » 22 Nov 2012, 03:06

Rares, this is an improvement. With daily data it is possible to determine enough cycles accurately to get a useful result I think.

I would still recommend using logs of data, and trying both the log price and rate of change of log price. See what cycles show up with each. I find:

For log price, with about 1000 cycles found we might expect just one at better than the p=0.001 level, but there are in fact 28 of them, and many more not far behind that. We can say that most of these cycles are real.

Code: Select all
no.cycles   period    date/time maximum    amplitude   Bartel's p
1218.95   6.25949   727568.480664322   0.000977   0.00004468
1126.95   6.77049   727568.665505873   0.000948   0.00011898
414.94   18.3882   727577.559420805   0.003052   0.00012837
283.74   26.89081   727576.678875882   0.003888   0.00015561
86.77   87.93362   727611.696478278   0.012519   0.00019019
589.9   12.9344   727573.291923706   0.002147   0.00019233
70.91   107.60119   727634.448343619   0.017107   0.00019509
87.96   86.74397   727620.382977504   0.012858   0.00020188
101.95   74.84061   727622.101938511   0.011905   0.00022728
353.06   21.61106   727580.281359645   0.003257   0.00027184
394.92   19.32037   727576.137090085   0.002674   0.00034635
1023.93   7.45168   727569.239866981   0.001   0.00036455
1177.88   6.47774   727567.993480346   0.000756   0.00037356
638.98   11.94091   727572.701553525   0.001756   0.00039561
557.94   13.67531   727572.352711589   0.00169   0.00046721
1239.91   6.15367   727568.174319735   0.000792   0.00049200
262.68   29.04675   727577.389436396   0.00396   0.00050597
178.18   42.82187   727603.825365328   0.005863   0.00051233
1202.95   6.34274   727568.598663597   0.000843   0.00053916
742.98   10.26946   727570.949030011   0.001279   0.00076347
176.99   43.10978   727599.687783569   0.00678   0.00080694
553.52   13.78451   727568.221368947   0.001583   0.00082389
706.85   10.79437   727570.658825084   0.001361   0.00084662
183.86   41.49897   727588.625967068   0.005587   0.00084964
64.06   119.10709   727660.552195673   0.015221   0.00086573
310.93   24.53928   727580.221115349   0.003302   0.00089417
29.07   262.46988   727790.729314951   0.029063   0.00091526
1263.75   6.03759   727567.014455662   0.000556   0.00095498


Redoing spectrum with delta log prices gives these cycles. Far fewer are below p=0.001. The big change is a cause for concern and I want to understand why before going on.

Code: Select all
no.cycles   period    date/time maximum    amplitude   Bartel's p
585.00   13.04103   727575.357886766   0.045141   0.00053209
962.06   7.92986   727572.801543096   0.044652   0.00054407
584.04   13.06246   727578.114153461   0.041696   0.00103445
294.01   25.9481   727591.858256466   0.044214   0.00114312
335.06   22.76906   727586.856395421   0.043518   0.00133766
101.72   75.00000   727598.973904927   0.042177   0.00232339
1227.33   6.21593   727566.680855726   0.040705   0.00236366
176.76   43.16022   727586.915962587   0.042429   0.002541
589.58   12.93972   727568.679827271   0.040321   0.00267618


One thing that worried me with this data was the sudden huge increase of over 20% in about September 1999. Is this real? Even if it is, such big changes can skew all the analysis because most statistics include the square of differences. Anyway, CATS has a means of reducing extreme values beyond some number of standard deviations, so I used this just to see what happened. With delta log prices gold:
( )3.0
This removes greater than 3 standard deviations by compressing outliers in a smooth manner. The spectrum of this:

Code: Select all
no.cycles   period    date/time maximum    amplitude   Bartel's p
962.04   7.93002   727572.828802206   0.047102   0.00034553
335.01   22.77246   727585.867661602   0.046221   0.00055099
583.94   13.0647   727576.842000101   0.044618   0.00055408
585.08   13.03924   727576.204101932   0.045693   0.0006344
1227.28   6.21619   727566.462707223   0.043666   0.00072933
101.68   75.0295   727597.769622087   0.042115   0.00128541
262.23   29.09278   727591.738408504   0.044366   0.00137268
414.61   18.40042   727570.265866398   0.041439   0.0019858
294.05   25.94457   727566.62010748   0.042233   0.00217768
598.14   12.75454   727576.339133531   0.040972   0.00231024
283.73   26.88824   727569.361982824   0.041202   0.00257107


We see some changes, but cycles 7.93, 22.77, 13.06, 13.04, 75.0 days are still there. These seem to be real.

Notice the two periods 13.06 and 13.04 days which are very close together, having only one cycle difference in the whole period of time. This means that the cycle is fading in and out of existence in a period of near 20 years. In the 2nd analysis above there are also cycles of 12.94, 13.68, 13.62, 12.75 and 12.43 days. This region clearly has something going on, but the many cycles periods indicates that we cannot use a single period throughout the 20 years because there are many modulations.

My suspicion would be that these are all related to half the solar rotation period, which itself will fluctuate somewhat due to magnetic changes in the Sun. So that period needs to be looked at on its own. I know that this all seems a distraction from a simple cycles analysis, but if you want to study markets, the fact is that anything simple is already known and already taken out of the markets to a large extent.

I did a cycles fit for those periods 7.93, 22.77, 13.06, 13.04, 75.0 days but it isn't very exciting. Cycles being significant doesn't necessarily mean that they are tradeable profitably. The combined correlation is only 0.1 which means perhaps making 55% profitable trades and 45% unprofitable. But your broker would get rich.
User avatar
RayTomes
 
Posts: 128
Joined: 02 Aug 2010, 23:24
Location: Auckland, New Zealand


Return to CATS Software

Who is online

Users browsing this forum: No registered users and 3 guests

cron