The ProRealAlgos proved algo development process

Chapter 2D is about trading hours, days and months. Here is what a real one looks like when you can check the answer.

We generated two price series. One has a genuine effect planted in it: the first four hours of each day carry real upward drift and the rest of the day is noise. The other is noise all the way through. The same simple rule, long inside a four hour window and flat outside it, was run on both.

Series with the planted effect: profit factor 3.44 over 104 trades. Series without it: profit factor 1.10 over the same 104 trades.

The gap is the whole point. A time filter that is picking up something real keeps working when you move the window a little. A time filter fitted to noise falls apart the moment you do, and on a single backtest the two look alike.

A strategy with seven adjustable numbers was run on a price series with no edge in it whatsoever. Not a weak edge. None. The series is a random walk and we know that because we generated it.

Tuning the seven numbers produced a rising equity curve over about a hundred trades. Then nothing was changed except the random seed, which produces a different stretch of the same edgeless process:

Seed 7, profit factor 1.09. Seed 11, 0.83. Seed 23, 1.34. Seed 44, 0.71. Seed 91, 1.24.

A profit factor of 1.34 on a hundred trades looks like a finding. It is nothing at all. That spread, on data guaranteed to contain no edge, is what overfitting produces, and it is the number your own backtest is competing with.

Five price series were generated. One has a specific rule planted in it: after three consecutive down closes, the next bar carries real upward drift. The others do not.

The matching strategy, buy after three red bars and hold five bars, was run on each, averaged over six seeds so that one lucky run could not decide it:

Planted edge, profit factor 1.84. Session effect, 1.55. Mean reverting, 1.10. Trending, 0.96. Pure noise, 0.86.

The planted one does stand out, over six runs. On any single run it does not: 1.84 and 1.10 overlap comfortably in one sample. That is the finding worth carrying around. One backtest cannot separate a real edge from a lucky one, and no amount of staring at the equity curve changes that.

ProRealTime will optimise your variables for you, and the number it hands back is almost always a good one. This is what that number is worth.

A moving average cross with two parameters, fifty six combinations in the grid. Three years to optimise on, then the next year traded with whichever pair won. Rolled forward six times, across sixty separate price series.

On series with no edge in them at all: best in-sample profit factor 1.62, and 0.62 in the year that followed. The chosen parameters beat 1.0 out of sample in 37 per cent of the 265 folds. Walk-forward efficiency, which is the out-of-sample result divided by the in-sample one, came out at 38 per cent.

Then the same test on series with a genuine effect planted in them: 1.71 in sample, 0.65 out of sample, above 1.0 in 41 per cent of folds. Four percentage points better than noise. A fifty six cell grid search could not find a real edge through its own optimisation noise.

This is why walk-forward exists, and also why it is not a rescue. It tells you honestly that the parameter you picked was worth about 38 per cent of what it looked like. It does not hand you a better parameter.

A strategy can have a real edge and still lose money. This is the experiment that shows it most cleanly, and it is the one most people never run.

A short horizon rule, three rising closes then hold five bars, on a series with a genuine effect planted in it. About 455 trades per run, average trade size 18 points, across forty separate series. Then the same thing again with a spread charged on every round trip:

No spread, profit factor 1.09. Half a point, 1.03. One point, 0.98. Two points, 0.88. Four points, 0.71.

The edge is real. We planted it. It is worth less than one point per trade, so a one point spread is the whole of it. The same run on a series with no edge went 0.99, 0.93, 0.88, 0.78, 0.61.

Two things follow. The first is to put your broker's real spread into the backtest before you believe anything, not after. The second is the harder one: the faster a strategy trades, the larger the share of its edge the spread takes, which is why short horizon systems look so good before costs and so ordinary after them.

Take a system that genuinely works and change nothing about it except how much you risk per trade. A 45 per cent win rate paying 1.8 to 1, which is plus 0.256 of a risk unit per trade, measured over 200,000 trades. Two hundred trades per path, fifty thousand paths at each setting.

Risking 0.5 per cent: median result 1.29 times the account, median worst drawdown 5 per cent. One per cent: 1.65 times, 10 per cent drawdown. Two per cent: 2.61 times, 19 per cent drawdown.

Five per cent: 8.28 times the account, median worst drawdown 43 per cent, and a 26 per cent chance of being down more than half at some point along the way. Ten per cent: 27.29 times, median drawdown 70 per cent, and the account halves at some point in 96 per cent of paths.

Read that last line again. The median outcome at ten per cent risk is twenty seven times your money, and almost every single path goes through losing half of it. That is the trap. The average keeps improving as you size up, long after the path has stopped being something a human being can sit through, and the drawdown you cannot sit through is the one you close at the bottom of.

The same profitable system. 45 per cent of trades win, the winners pay 1.8 times the losers, and over 200,000 trades it makes 0.256 of a risk unit each time. It works. Here is what working looks like from the inside.

Two hundred trades, fifty thousand simulated runs. The median worst losing streak is eight trades in a row. The ninetieth percentile is eleven. The ninety ninth is fifteen. The longest run seen in fifty thousand was thirty two.

A run of six or more happens in 92.9 per cent of those lives. Eight or more in 53.1 per cent. Ten or more in 20.0 per cent. Twelve or more in 6.4 per cent.

So eight losses in a row is the normal experience of a system that works, not evidence against it. If your plan is to stop after five, you have written a plan that turns off nearly every working system you will ever build.

A system you licensed is in drawdown. You have what you would really have: the record you were shown before you bought it, and the record since. Keep it running, or stop it?

Three worlds were generated, three hundred cases each, every one of them showing a backtest profit factor of about 1.28. In the first the edge is still there. In the second it dies the day you go live. In the third there never was one and the backtest was a good run on noise.

Over the first hundred live trades: edge still there, median live profit factor 1.14, above 1.0 in 69 per cent of cases. Edge died at go-live, 0.99, above 1.0 in 49 per cent. Never had an edge, 1.02, above 1.0 in 53 per cent.

Now the number that matters. Hand someone one live record from a healthy system and one from a dead one and ask which is which. Picking the higher live profit factor is right 64 per cent of the time. Against a system that never worked at all, 62 per cent. Guessing is 50.

A hundred real trades buys you fourteen points of edge over a coin toss. That is the honest answer, and it is why what saves you is not a cleverer diagnosis. It is position size, and a rule you wrote down before you started.

Decide this while nothing is at stake, because you will not decide it well halfway down. A usable rule names a number and a horizon, not a feeling.

Three of the pieces are already in the ProRealAlgos support answers: run the systems together, because one is in drawdown while another is making money; check in weekly or monthly rather than daily; and remember that stopping an algo does not close its open positions.

The two lessons above give you the numbers to put in it. Eight losses in a row is ordinary. A hundred trades barely tells a dead system from an unlucky one.

The last gap between a backtest and a running system is a short list of platform facts, and every one of them has cost somebody money.

Stopping an algo does not close what it already has open. The position stays, without the code that was managing it. If you want to be flat, close the position.

A backtest fills at the bar close. Live, you fill at whatever the next tick is, and on an instrument with a two point spread that difference is the whole of a short horizon edge, as the cost lesson above measured.

A strategy that has been restarted begins with no memory. Counters, flags and trailing stop levels held in variables are all back to their starting values, so the first bar after a restart is the one to write defensively.

Run it on a demo account for long enough to watch it take, manage and close a trade unattended. Not to prove the edge, which a demo cannot do, but to prove the plumbing.

Every page of the reference ends with a section called common errors and gotchas. There are 1,168 of them across 301 instructions, each one a specific way that correct looking code does the wrong thing.

They are also the shortest way to check code you did not write. An AI will happily hand you ADX[14](close), which looks reasonable and is a syntax error in ProRealTime, and the ADX page says so in one line.

Open the reference