Forward Testing on Demo
The simulator is passed, the exams are stitched. Now: learner plates. Real roads, real traffic, fake stakes — the last proving ground before real money, and the one where most illusions die.
11.1Why the road test exists
A backtest — even a walk-forward one — is still driving on recorded roads. puts the strategy on today's live roads with learner plates: a demo account, where prices, spreads and news are completely real and only the money is pretend. It catches what no recording can:
Real fills. Backtests assume you get the price you saw. Live, there's — usually pennies, occasionally decisive, always absent from recordings.
Real spreads. Recorded spreads are averages; live spreads breathe — tight in the London tide, fat at 5am, gaping around news.
Real infrastructure. Does the scheduler fire? Does MT5 stay connected overnight? Does the work? Software problems only show up when software actually runs.
Real you. The strategy is tested — but so is the pilot. Watching a drawdown happen live, with your ritual holding, is training no simulator provides.
11.2How IronGate runs the road test
Two lanes again, both safe by construction:
Paper mode — the follows live MT5 prices and tracks hypothetical trades without ever placing an order (its code physically contains no order calls — the steering wheel isn't connected). You start paper sessions per pair from the backtest UI and watch how a backtested parameter set performs on live data. Demo mode — the full pipeline places actual orders on the demo account: scheduler wakes the recommender, cards appear, you accept or reject, orders fill, the journal records. The complete machine, rehearsing end-to-end:
11.3The hard part — judging small numbers honestly
Here's where beginners fool themselves in both directions. Twenty forward trades come in at a 35% win rate when the backtest promised 45% — disaster? Or 60% — genius? Usually: neither. Small samples wobble, hugely. The question is never "does forward match backtest exactly?" but "is the difference bigger than luck can explain?"
⚙ The wobble lab — is the drift real or just luck?
Try 7 wins from 20 (feels awful, is normal), then 200 trades with the same 35% rate (now it's a real signal). The lesson: trade count is what turns feelings into evidence — the same principle as Lesson 8's "hundreds, not dozens."
Give it 30+ trades or 4–8 weeks before judging performance at all. Before that, judge only process: did every trade follow the ritual? Did the guards fire when they should?
Measure in R (Lesson 2), never dollars — and compare expectancy, win rate, and drawdown against the backtest's, expecting wobble.
When drift is real (outside the wobble band with decent sample size), the suspects in order: costs harsher than modelled → regime changed → a bug in the pipeline → the backtest was overfit after all. Each has a different fix; none of them is "abandon everything in a panic."
11.4Your forward-test protocol
Start in paper mode on 2–3 pairs the backtests favour. Zero risk, live data, watch the machine think.
Graduate to demo orders with the full approval loop — you accepting and rejecting cards using the three-question ritual (Lesson 3) plus the regime match check (Lesson 5).
Weekly review against backtest expectations, in R, with the wobble lab's honesty about sample size. The journal (next lesson) is where this happens.
Months, not days. The demo phase ends when you have enough trades to trust the comparison and you've stopped being surprised by your own system. There is no prize for rushing to live — the market will still be there.