The only thing I did last week was to get rid of the short silver position. Between that and the fluctuations of my NZD/USD long position the account was down 1.3% for the week.
I'm still in the second phase of my big 3 part research project, looking at reversal patterns now. After 12 more pages of research notes, the best thing I've come up with so far is that when there are two or more consecutive bars with equal or higher highs and the last bar's close is at least 40% of that bar's range away from its high, there's about a 20% chance that we're seeing a significant top. The reverse is true for bottoms. This is for daily bars on EUR/JPY only.
I'm using the statistical concepts of type-I and type-II error (false positives and false negatives) in this search for indicators of bottoms and tops. Even if I don't find any mind blowing predictive patterns, at least I've developed a good new research tool and have a better understanding of what these kind of searches are all about.
To give some flavor for the kind of information I'm able to generate, the topping signal that I described above occurs about 150 times in my dataset of 1400 bars. There are about 70 significant tops in the data, so they happen about 5% of the time (I'm rounding all these numbers off for clarity). Of the 150 signals, 20% are correct, so the signal found about 30 of the tops and missed 40 of them.
So we have this breakdown:
Top present? Yes No
Signal 30 120
No signal 40 1210
False positives = 30/150 = 80% Hit rate = 20%
False negatives = 40/70 = 57% Capture rate = 43%
All of this points out a common tradeoff in trying to find any kind of signal (not just in trading but in medical tests and all other kinds of testing). If you make the test too broad (i.e. a signal exists any time a price bar happens), the capture rate will be very high (low type-II error) but hit rate will be very low (high type-I error). If on the other hand you make the signal too specific (only if conditions A, B, C, D, and E are met), then the hit rate might be high (low type-I error) but the capture rate will be low (high type-II error) meaning that a lot of opportunities will be lost due to the strict signal rules.
Anyway, the research continues. I may try some simple "which way is the market going now" type trades this week, but I don't really have a bread and butter method yet.
I'm still in the second phase of my big 3 part research project, looking at reversal patterns now. After 12 more pages of research notes, the best thing I've come up with so far is that when there are two or more consecutive bars with equal or higher highs and the last bar's close is at least 40% of that bar's range away from its high, there's about a 20% chance that we're seeing a significant top. The reverse is true for bottoms. This is for daily bars on EUR/JPY only.
I'm using the statistical concepts of type-I and type-II error (false positives and false negatives) in this search for indicators of bottoms and tops. Even if I don't find any mind blowing predictive patterns, at least I've developed a good new research tool and have a better understanding of what these kind of searches are all about.
To give some flavor for the kind of information I'm able to generate, the topping signal that I described above occurs about 150 times in my dataset of 1400 bars. There are about 70 significant tops in the data, so they happen about 5% of the time (I'm rounding all these numbers off for clarity). Of the 150 signals, 20% are correct, so the signal found about 30 of the tops and missed 40 of them.
So we have this breakdown:
Top present? Yes No
Signal 30 120
No signal 40 1210
False positives = 30/150 = 80% Hit rate = 20%
False negatives = 40/70 = 57% Capture rate = 43%
All of this points out a common tradeoff in trying to find any kind of signal (not just in trading but in medical tests and all other kinds of testing). If you make the test too broad (i.e. a signal exists any time a price bar happens), the capture rate will be very high (low type-II error) but hit rate will be very low (high type-I error). If on the other hand you make the signal too specific (only if conditions A, B, C, D, and E are met), then the hit rate might be high (low type-I error) but the capture rate will be low (high type-II error) meaning that a lot of opportunities will be lost due to the strict signal rules.
Anyway, the research continues. I may try some simple "which way is the market going now" type trades this week, but I don't really have a bread and butter method yet.