Validation & Deployment

Split-half is a good first filter. Walk-forward is the honest version.

Optimise on a window (say 12 months), then trade the next window (say 3 months) with those frozen parameters. Roll both windows forward and repeat. Stitch together only the out-of-sample segments — that curve is the closest thing to a truthful simulation of having actually run the strategy.

What you are testing is not just the strategy but the process of maintaining it. If re-optimising every quarter produces wildly different parameters, the strategy is unstable even if the equity curve looks fine.

Watch for walk-forward efficiency: out-of-sample performance divided by in-sample. Below about 0.5 means most of your backtest profit was fitting, not edge.

The takeaway: if a strategy needs frequent re-optimisation to keep working, you do not have a trading system. You have a curve-fitting habit with a P&L attached.

Once a strategy passes, your job is to attack it. Every test below is one we run before anything ships.

Cost stress. Double the spread. Does the edge survive? A thin edge that dies at 1.5× costs will die in live trading.

Parameter neighbourhood. Vary every parameter ±20%. Performance should degrade gracefully. A cliff edge means you found a coincidence.

Instrument transfer. Does it work on related instruments? If a "universal" strategy only works on one pair, be suspicious — although the reverse also matters: our Straddle works on USD/JPY and gold but is negative on EUR/USD, GBP/USD and USD/CHF, and we publish that because knowing where it fails is part of knowing what it is.

Randomised entries. Replace your signal with a coin flip, keep the exits and sizing. If results barely change, your "edge" was in the risk management, not the signal.

The takeaway: you are not trying to confirm the strategy. You are trying to kill it. If you cannot, it may be worth trading.

Meta-labelling separates two questions that beginners conflate: which direction? and should I take this trade at all?

The primary model (often a simple rule) decides direction. A secondary classifier then predicts whether that specific signal is likely to be a winner, using features the rule ignores — volatility regime, session, spread conditions, recent behaviour. You trade only the signals the second model approves.

The appeal is that a binary "take it or skip it" problem is easier to learn than direction, and it can cut drawdown sharply by removing the worst trades.

The honest limits matter. Our own meta-label model is trained on 331 labelled entries with a validation accuracy of 0.663. That is a small sample, and accuracy is not profitability — a model can be accurate on the easy cases and wrong on the ones that cost money. We treat it as research, not as a finished product, and we say so.

The takeaway: meta-labelling is a filter, not an oracle. It can only remove bad trades from an edge that already exists — it cannot create one.

Backtests execute perfectly. Live trading does not, and the gap is where thin edges die.

Slippage is worst exactly when you need it least — at breakouts, on news, at session opens. A strategy whose edge is +0.06R per trade can be wiped out by half a pip of consistent slippage.

Latency matters more than people think for stop orders: your platform, VPS, connection and broker each add delay between signal and fill.

Partial fills and requotes break the assumption that you got the whole position at one price. Weekend gaps mean your stop is a request, not a guarantee.

This is why demo-forward testing is not optional. Run the strategy live on a demo account for weeks, then compare the fills you actually got against what the backtest assumed. That comparison, not the backtest, is what tells you the edge is real.

The takeaway: the backtest tells you whether the idea has merit. Only forward testing tells you whether you can actually capture it.

You have an edge that survived walk-forward, the nasty test and demo-forward. Now scale in deliberately.

Start at micro size — smaller than the maths justifies. The first live weeks are not about profit, they are about verifying that your live results track your tested expectancy. Set a rule in advance: if live expectancy is materially below tested after N trades, you stop and investigate.

Then audit relentlessly. After every trade, classify it on two axes: was the process correct, and was the outcome good? That gives four boxes. A losing trade taken correctly is a good trade. A winning trade taken by breaking your rules is the most dangerous outcome in trading, because it teaches you to break rules.

Track your edge in R over rolling windows. Real edges have flat stretches that feel exactly like being broken. Deciding in advance what "broken" looks like — a number, not a feeling — is what stops you abandoning a working system during normal drawdown.

The takeaway: judge yourself on process, not P&L. The market controls the outcome of any single trade. You control whether you followed the plan.