stevehopwoodforex.com
https://www.stevehopwoodforex.com/phpBB3/
Print view

Daily Fibonacci Signals
https://www.stevehopwoodforex.com/phpBB3/viewtopic.php?t=4100		 Page 8 of 21
Author:  snailbeard [ Wed Apr 06, 2016 9:15 am ]
Post subject:  EODGT version
After reimplementing the end-of-day guess and running back tests I found that version 2, still works well on Rio and Gold but also now works on Barc. 3I did best with a profit factor of 3. It was getting close to the end of day but I went through and replaced version 1 with 2 on several charts. Rio and Gold are running live at the moment with reduced risk. From the following chart you can see that the opening guess was correct but momentum is weak and price might not reach the initial target. There is currently no break-even or profit-lock.
Rio&Gold-live-trades-6Apr2016.png
Gold is currently moving in the right direction, but not with any significant momentum. Price is getting strong support at significant Fibonacci levels around 1225.
sell-gold-but-strong-Fib-Support-at-1225-6Apr2016.png
Author:  snailbeard [ Wed Apr 06, 2016 3:17 pm ]
Post subject:  Bearish engulfing & RSI > 70
Whenever a trade hits a stop-loss because it failed to reach the target we are tempted to rush in with break-even and profit-locking functions. However, sometimes these have an unexpected long term impact. We need a statistically meaningful number of trades to have any hope of proving one way or the other whether or not it has any benefit. I suspected that Rio would hit the stop-loss having missed the target this morning. I'm quite sure from previous tests that a simple single indicator based exit would be detrimental. However, I would like to try combining a number of events into a composite decision. For example, in the following chart not only does RSI indicate over-bought but we also get a bearish candle stick pattern at the same time:
RioTinto-RSI&Bearish-Engulfing-for-smart-exit-6Apr2016.png
Author:  snailbeard [ Thu Apr 07, 2016 8:48 am ]
Post subject:  More accurate back testing
Those of us who make use of MetaQuotes CrappyTester are familiar with it's limitations. We will usually get a rose-tinted view of results. This is especially true, when gaps are involved. A stop-loss in the middle of a gap does not get triggered in real life, but CrappyTester ignores gaps and gives you the price that you asked for. The implications are that you either need a better P/L ratio or a better win to loss ratio (or a different approach).

The following chart shows the difference between the ATR derived stop-loss price and the actual exit price. Perhaps it doesn't make sense to use a passive stop-loss, or the passive stop-loss should be larger than the ATR.
riotinto-gapping-undermines-stop-loss-7Apr2016.png
So you might wondering why a sell entry was predicted when all the moving averages are going up. That's because often the morning gap is in the opposite direction to the price direction, otherwise we would all be rich!
Author:  snailbeard [ Thu Apr 07, 2016 4:29 pm ]
Post subject:  Part 2 - More accurate back testing
After replacing the hard stop losses with an actively monitored stop-loss the results are quite different.
Crappytester is relieved of it's stop-loss monitoring function and you could call it a stealth stop-loss facility.

The profit factor has been reduced from a rosy 2.3 to a more typical 1.5. Too many wrong gaps and profits will quickly turn into losses because a bad gap is equivalent to a sequence of 4 losing trades according to the nominal stop-loss.

The following extract shows the cost of a bad guess when the gap is against us:

Code: Select all

35	2016.03.07 18:15	buy__	18	210	22.52	20.474	23.02	00000	5458.05
36	2016.03.08 10:01	close	18	210	21.43	20.474	23.02	-228.9	5229.15
[/size]

And the reduced profit curve is now more lumpy:
Results-after-compensating-for-crappertester-sl-errors-07Apr2016.png
Author:  snailbeard [ Tue Apr 12, 2016 7:41 am ]
Post subject:  overcoming curve fitting
All my trading algorithms are looking for patterns which I believe give a better chance of success. Experienced traders know that these patterns are constantly changing over time.

So is it possible for an EA tuned to Riotinto to be used for XAUUSD without any changes?
Lets take it a step further and trade Riotinto on FxPro and Gold on Global Prime with the same EA and the same filters :lol:

Now we all think from prior experience with EAs and brokers that just changing the broker is enough to mess up an EA. It doesn't seem to make any sense, firstly one asset only trades a few hours per day and often starts the following day at some significantly different price , whereas gold trades 24x7 and doesn't generally have big gaps.

After trying out lots of tweaks and variations I rolled back to a version with fewer filters and these are the result on FxPro with most filters off:
riotinto_D1_3M1_no_filters_11Apr2016.png
Now with a bit of extra tuning, which might amount to curve fitting (unless it works on other assets), I can avoid certain patterns leading to bad trades. So these following results and settings are the ones that need to be proven against other price data because the results are suspiciously favourable:
riotinto_with_best_filters_3M1_11Apr2016.png
Another question to ask is:
What is the chance of gold trades resulting in a higher profit factor than Riotinto (2.9)?
Looking at the xauusd results first to see if it is worth going any further, how does this compare to our prior expectations?
gold-trade-graph-profts-GP-Q1-2016-11Apr2016.png
Gold has a profit factor just marginally better than Rio, it is difficult to believe, perhaps it is an anomaly.
However, we can better understand the EAs performance by studying where it makes more losses.

The first part of the graph is lossy and most profits are made after the initial lossy phase, so relating this to the price chart:
EODGT-xauusd-trend-losses-range-profits-Q1-2016.png
The xauusd data covers the same period but is visibly very different, the first part has an up trend and the second part goes into ranging, so we don't need to be concerned that the data is skewed in favour of the EA.
This is still too little data to be sure that we can say trends are lossy and ranging is profitable, but it would be good to have EA which works well for ranging assets.

So I am wondering if can be adapted for other pairs...
Author:  snailbeard [ Tue Apr 12, 2016 11:37 am ]
Post subject:  Trend filtering
This is the first time I have thought about suppressing trading during a trend, all past efforts have been towards working out whether or not price is trending and on some time scale there has to be a trend to follow.

It turns out that, for Gold, D1 MACD currently makes a good directional indicator, so I have added an optional filter to prevent counter trading while there is strength in the trend. I was concerned that this could end up filtering out profitable trades as well, but the filtered entries are all in first zone:

Code: Select all

2016.01.14 18:15  EODGT_v2e XAUUSD,M1: D1TrendingUp  bGuessFalling set false
2016.01.18 18:15  EODGT_v2e XAUUSD,M1: D1TrendingUp  bGuessFalling set false
2016.01.25 18:15  EODGT_v2e XAUUSD,M1: D1TrendingUp  bGuessFalling set false
2016.02.05 18:15  EODGT_v2e XAUUSD,M1: D1TrendingUp  bGuessFalling set false
2016.02.11 18:15  EODGT_v2e XAUUSD,M1: D1TrendingUp  bGuessFalling set false
[/size]

That's enough tweaking for now, it's time to do some more live forward testing.

Although, it is tempting to add a switch to do trend trading when D1 is trending...
Author:  snailbeard [ Mon Apr 18, 2016 11:32 am ]
Post subject:  Best (trending) entry method for gold
While the forward test on xauusd range trading continues, I spent some time experimenting with several entry methods for gold. Enter on higher-highs or lower-lows is a popular method with shares and I have seen long term profits with this method even on some currency pairs. The down side is that we sometimes get a spike in the high or low which has a knock on effect by delaying the next entry. This in turn reduces the available remaining swing of the day, which means either running the trade for longer (increased risk of reversal) or taking a reduced profit.

Using D1 MACD to filter out ranging periods I had the following results for enter by HH or LL:

Code: Select all

1	2016.01.25 09:40	buy	1	0.12	1104.66	1073.07	1136.25	0	5000
2	2016.01.26 05:30	close	1	0.12	1112.79	1073.07	1136.25	68.41	5068.41
3	2016.01.26 05:30	buy	2	0.12	1112.99	1081.05	1144.93	0	5068.41
4	2016.01.26 20:50	close	2	0.12	1120.68	1081.05	1144.93	64.97	5133.38
5	2016.01.27 21:45	buy	3	0.13	1124.95	1093.62	1156.28	0	5133.38
6	2016.01.28 03:05	close	3	0.13	1120.54	1093.62	1156.28	-41.28	5092.09
7	2016.02.01 04:15	buy	4	0.13	1121.47	1092.53	1150.41	0	5092.09
8	2016.02.01 22:55	close	4	0.13	1129.14	1092.53	1150.41	70.2	5162.29
9	2016.02.02 04:35	buy	5	0.13	1127.45	1098.74	1156.16	0	5162.29
10	2016.02.02 19:20	close	5	0.13	1122.77	1098.74	1156.16	-42.83	5119.46
11	2016.02.03 15:25	buy	6	0.13	1132.88	1105.16	1160.6	0	5119.46
12	2016.02.03 15:40	close	6	0.13	1128.51	1105.16	1160.6	-40	5079.46
13	2016.02.04 11:05	buy	7	0.13	1146.07	1118.76	1173.38	0	5079.46
14	2016.02.04 15:50	close	7	0.13	1154.26	1118.76	1173.38	74.96	5154.42
15	2016.02.08 14:35	buy	8	0.13	1177.96	1147.75	1208.17	0	5154.42
16	2016.02.08 16:25	close	8	0.13	1185.92	1147.75	1208.17	72.86	5227.28
17	2016.02.11 02:45	buy	9	0.13	1213.49	1180.52	1246.46	0	5227.28
18	2016.02.11 02:55	close	9	0.13	1208.61	1180.52	1246.46	-44.67	5182.61
[/size]

In response, I thought of trying another common entry method based on price action. Now using the previous D1 Close as the reference price, we can measure price swing in the D1 trend direction, wait for price to move enough, I also look for a pullback followed by some good shorter term price action momentum to confirm the trend.

This works well for the recent data, but the method is possibly curve fitted to the data, there needs to be additional parameters to make it into an companion to manual trading. However, the price action method is significantly more profitable:

Code: Select all

1	2016.01.25 05:20	buy	1	0.12	1102.11	1070.52	1133.7	0	5000
2	2016.01.25 22:40	close	1	0.12	1107.74	1070.52	1133.7	47.61	5047.61
3	2016.01.26 05:30	buy	2	0.12	1112.99	1081.05	1144.93	0	5047.61
4	2016.01.26 20:50	close	2	0.12	1120.68	1081.05	1144.93	65.03	5112.64
5	2016.02.01 08:00	buy	3	0.13	1122.29	1093.35	1151.23	0	5112.64
6	2016.02.01 23:35	close	3	0.13	1128.69	1093.35	1151.23	58.63	5171.27
7	2016.02.03 15:25	buy	4	0.13	1132.88	1105.16	1160.6	0	5171.27
8	2016.02.03 15:40	close	4	0.13	1128.51	1105.16	1160.6	-40.03	5131.24
9	2016.02.04 11:05	buy	5	0.13	1146.07	1118.76	1173.38	0	5131.24
10	2016.02.04 15:50	close	5	0.13	1154.26	1118.76	1173.38	75.03	5206.27
11	2016.02.08 14:35	buy	6	0.13	1177.96	1147.75	1208.17	0	5206.27
12	2016.02.08 16:25	close	6	0.13	1185.92	1147.75	1208.17	72.92	5279.19
13	2016.02.09 05:45	buy	7	0.13	1193.65	1159.94	1227.36	0	5279.19
14	2016.02.09 09:45	close	7	0.13	1189.33	1159.94	1227.36	-39.57	5239.62
15	2016.02.10 22:20	buy	8	0.13	1193.74	1161.36	1226.12	0	5239.62
16	2016.02.10 23:20	close	8	0.13	1197.09	1161.36	1226.12	30.69	5270.31
17	2016.02.11 02:15	buy	9	0.13	1205.24	1172.27	1238.21	0	5270.31
18	2016.02.11 02:45	close	9	0.13	1213.29	1172.27	1238.21	73.74	5344.05
[/size]

It might also be possible to tweak the stop-loss to take-profit ratio.
Author:  snailbeard [ Thu Apr 28, 2016 3:23 pm ]
Post subject:  Stoploss manipulation
In a previous post I reported that CrappyTester ignores gaps and so can give false prices for stop-loss or take profit. I have just tweaked the code to get round this issue. The solution for CrappyTester is not good when running live because the long distance oversized stop-loss increases the risk on a live account of a much bigger loss since it relies on the EA being constantly running without interruption, in order that the EA closes the trade at the stealth stop-loss.

One solution is to make sure that the EA reload the stealth stop-loss when it is restarted, but what if it does not get restarted?

So what are we trying to achieve by manipulating the stop-loss?
1. To get more realistic exits from CrappyTester.
2. To hide our real SL/TP from the stop hunters.

So having revisited the code it seems prudent when not running in CrappyTester to use a much smaller stop-loss offset in case price suddenly moves while the EA is not running. I am not sure what this live-offset should be because it depends on how far the crims can manipulate the price, so I am currently getting the (average spread x 2) and using that. The nominal spread for gold is 30 points and this frequently hits 40 on FxPro.

So this how the same trade should look when run first in Test mode then live

Code: Select all

Pre-entry SL/TP values
---------------------
dAdjustedLots: 0.09 
dTrigger: 1212.28
rTakeProfit: 1223.88 
rStopLoss: 1206.48

Hard super oversized SL/TP values for CrappyTester
---------------------
open #1 buy 0.09 GOLD 
at 1212.28 
sl: 1159.89 -1206.48 = −46.59
tp: 1270.47 -1223.88 = 46.59

Prc: 1212.280000, 
M.SL: 1206.480000, 
M.TP: 1223.880000 
 
========================
Still CrappyTester but now pretending to run live with an slightly oversized hard stop-loss
 
dAdjustedLots: 0.09 
dTrigger: 1212.28 
rTakeProfit: 1223.88 
rStopLoss: 1206.48

Actual hard slightly oversized SL/TP values for live trading:
---------------------
open #1 buy 0.09 GOLD 
at 1212.28 
sl: 1205.88 -1206.48 = −0.60 (twice the nominal spread)
tp: 1224.48 -1223.88 = 0.60
[/size]

Its hard to know what it will actually do when running live, except by running it live...
Author:  snailbeard [ Sat Apr 30, 2016 7:12 am ]
Post subject:  Gold ranging issues
Just when spring was looking so good came a series of April showers. The first half of April started very well but in the second half of April something was different. After looking carefully not only is the D1 ATR very low but H4 shows a cyclical swing across a very flat H4 moving averaging and perhaps almost perfect conditions for stochastic based overbought oversold entries. However, it messes up EODGT which is a short-term-trend predictor.

In the following chart we can see how some of the trades are in the wrong direction because of the H4 price oscillation:
Gold-H4-Range-Trading-Losses-April2016.png
So we need a filter which understands this but doesn't interfere with the successful trades, the aim being to keep the win to loss ratio at about 70% and the P/L ratio about 2.

So how flat is flat and how far on the wrong side of a flat MA is too far from the MA line to be a good entry?
One obvious choice would be to use the Bollinger boundaries based on standard deviation.
However, I have used a fraction of the average daily range instead of the Bollinger values.
Flatness is quite a difficult value to pin down, but again I have related this to the average daily range.

This filter should only be active if price is oscillating across a flatish line on a H4 timeframe.

This method so far eliminates 3 out of 4 wrong direction entries in April but at the moment kills too many earlier winning trades.
Author:  snailbeard [ Mon May 02, 2016 6:53 am ]
Post subject:  Better ranging analysis
After tweaking the H4 flatness detection there is an improvement in the selectivity of the April ranging filter, but this filter works much better when combined with D1 Ma40 when the slope is almost horizontal. So now it should be possible to add a third type of entry method specifically for this kind of ranging. It might be possible to use the H4 stochastic oscillator to help select an entry point.

In the following chart it is clear how flat the second part of April is before a higher-high trend break-out at the end of the month:
Gold-D1-FlatRangingFilter-April2016.png
All times are UTC Page 8 of 21

Powered by phpBB® Forum Software © phpBB Limited