Why your backtest lies

We analysed 17 robots that shone on paper. None with a real edge survived the market's ticks. With data, not opinions.

The trap that blows up accounts

Almost every robot (EA) bought and sold comes with a spectacular backtest: high profit factor, flawless equity curve. The catch is that this backtest uses a synthetic tick model —prices generated by the simulator— not the reality of the market: real spreads, slippage and the microstructure of every tick.

The consequence: a robot can boast a Profit Factor of 1.68 in the simulator and lose money in real execution. This is not theory. We measured it, robot by robot, with the broker's real ticks.

The proof: simulator vs. real ticks

We put a family of 17 robots (variants of the same gold idea: order-block entry plus a trend filter) through a double judge —survival on real ticks vs. the synthetic backtest. In the simulator several looked like pure gold. On real ticks, the truth (representative sample):

RobotSim PFReal PFReal net (1,000 $)
PHALANX1.680.74−878 $
TITAN1.460.84−571 $
IRON DOME1.400.89−395 $
ADAPTIVE1.330.78−820 $
VANGUARD1.250.74−196 $
GHOST PROTOCOL 21.010.68−775 $
V3.1 «Golden Phalanx»1.010.38−959 $
V3 «Oracle of Delphi»0.950.50−857 $

The pattern is uniform and damning: the synthetic mirage collapses against real fills. And note: the two «favourites» with a Kill Zone time filter (Golden Phalanx and Oracle of Delphi) were among the worst in reality —PF 0.38 and 0.50, probability of ruin near 100%. A pretty filter won't save logic with no edge.

The subtle case: surviving ticks isn't enough

One robot did survive real ticks (survival 103%). It looked like the good one. But a tail Monte Carlo —the worst case derived from its own logic, not the average— revealed a 41% probability of ruin. It's an «insurance seller»: it collects small premiums for months until a gap with the grid loaded wipes out the account. A 3-week backtest never sees that day.

How to validate for real (the four filters almost nobody applies)

  1. Real-tick survival: the same robot on real ticks vs. synthetic. If the result changes, it was a mirage.
  2. Tail Monte Carlo: probability of ruin with the exact tail derived from the robot's logic, not an optimistic average.
  3. Walk-forward: does the edge hold on periods the robot never saw while being tuned?
  4. Plateau (sensitivity): perturb the parameters ±15%. If a single tweak breaks it, it was overfitting in disguise.
What one that DOES pass looks like

For contrast, a robot bred and filtered with this process: out-of-sample PF 1.31 · real-tick survival 99.8% · probability of ruin 0% · plateau 6/6 healthy variants. The difference with the table above isn't luck: it's the method.

Conclusion

Don't trust a pretty backtest. Before putting a real euro into a robot —yours or bought— the right question is: does it survive real ticks, its tail risk, and perturbing its parameters? If you haven't measured it, you don't know.