Well, let us take a closer look at the results. The picture looks pretty, but what do the numbers tell us?

Well, I guess the PROFITS part is clear. The Profit Factor is 1.07, with a respectable Sharpe Ratio of 1.04 and a R2 of almost one. Pretty good for any strategy, lots of profits with not that much risk. The draw-downs don’t look terrible either, so Zorro Trader and DeepNet must really know what they are doing, as the percent of winning trades is 56.1%. Here are the detailed results:

WFA Test DeepLearn EUR/USD

Simulated account   AssetsFix 
Bar period          1 hour (avg 87 min)
Total processed     53972 bars
Simulation period   2014-12-05..2021-01-25 (37068 bars)
Test period         2015-12-16..2021-01-25 (30853 bars)
Lookback period     100 bars (6 days)
WFO test cycles     51 x 605 bars (5 weeks)
Training cycles     52 x 5445 bars (47 weeks)
Montecarlo cycles   200
Simulation mode     Realistic (slippage 0.0 sec)
Avg bar             12.3 pips range
Spread              0.0 pips (roll 0.00/0.00)
Contracts per lot   1000.0

Gross win/loss      16126$ / -15111$ (+10155p)
Average profit      199$/year, 16.55$/month, 0.76$/day
Max drawdown        -198$ 19.5% (MAE -198$ 19.5%)
Total down time     96% (TAE 96%)
Max down time       27 weeks from Aug 2019
Max open margin     50.18$
Max open risk       50.18$
Trade volume        35444350$ (6933371$/year)
Transaction costs   0$ spr, 0$ slp, 0$ rol
Capital required    202$

Number of trades    31170 (6098/year, 117/week, 24/day)
Percent winning     56.1%
Max win/loss        19.61$ / -31.40$
Avg trade profit    0.0326$ 0.3p (+9.2p / -11.0p)
Avg trade slippage  0$ 0.0p (+0.0p / -0.0p)
Avg trade bars      2 (+1 / -3)
Max trade bars      4 (5 hours)
Time in market      256%
Max open trades     3
Max loss streak     22 (uncorrelated 13)

Annual return       99%
Profit factor       1.07 (PRR 1.05)
Sharpe ratio        1.04 (Sortino 1.02)
Kelly criterion     1.09
Annualized StdDev   94.91% 
R2 coefficient      0.952
Ulcer index         5.2%
Scholz tax          268 EUR

Year Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Total
2015                                              -3    -3
2016  32  -7  15   8  32  -4  21  10   8  14 -10  -5  +116
2017 -22  10  41  -5 -12  55  -8 -21  23  -5  32  -5   +84
2018 -15  12  21  32  16  -8  -5 -14  -2  -1  -3   8   +42
2019  15  37  50   6  27  33  15  15 -26  24  -4 -20  +173
2020  -2  15   3  31 -13  12 -15   5  12  -8  11  46   +97

Confidence level     AR   DDMax  Capital
 10%                104%   185   192$
 20%                 97%   202   205$
 30%                 91%   219   218$
 40%                 85%   239   233$
 50%                 81%   256   246$
 60%                 75%   278   263$
 70%                 70%   305   284$
 80%                 62%   351   319$
 90%                 50%   449   394$
 95%                 46%   504   436$
100%                 31%   764   635$

Portfolio analysis  OptF  ProF  Win/Loss   Wgt%  Cycles

EUR/USD             .155  1.07 17475/13695  100.0  //XXXXXXX\XXXXXXXXX/XX//X\\XXX//
EUR/USD:L           .169  1.08 8515/7450   56.3  ///XX\//X\XX////XXX//\//\\\\\\//
EUR/USD:S           .140  1.06 8960/6245   43.7  //X//XXX/\X/X\\X\///\////\\/////

Now here’s what bothers me, and it bothers me A LOT. Although DeepNet seems to have learned something about predicting the stock market (56.1% is better than random chance) this little important line might be a killer:

Avg trade profit    0.0326$ 0.3p (+9.2p / -11.0p)

What this means is that although more than 56% of our trades are WINNERS, we only make about 3 cents per trade. But we do trade a lot, so these little cents add up to 99% annually. If only trading were free… But it’s NOT. There are broker commissions (more or less hidden), slippage, the bid/ask spread, etc. Let’s comment out this line in the script, like this:

//Spread = RollLong = RollShort = Commission = Slippage = 0;

… save and run the back-test again. No re-training of the DeepNet, just save and hit the Zorro Trader Test button again. What this basically does is it’s loading the market friction and cost variables from the AssetFix file, instead of making them 0. Here’s the result:

But wait, this is impossible! How come we now lost not only the 99% in profits but all our initial capital as well? The trades are exactly the same, Zorro Trader works the same, and DeepNet is still 56% right on predicting the future! Yes, but look at this line:

Transaction costs   -4676$ spr, 26.16$ slp, -15.44$ rol, -3.12$ com

Which makes the important line from before look now like this:

Avg trade profit    -0.12$ -1.2p (+8.2p / -10.6p)

Practically, our Broker (who charges the spread and the commission) and the Market Gods (who create the slippage) ate our lunch. We simply do not make enough money per trade to survive these costs, so no matter how clever a technology our Neural Network is, this strategy is not tradable. So, please, do not trade it at home! But I guess this is a great example of a nice strategy that does great in a no-costs back-test, but does not survive contact with reality.