Each stage answers a question the previous one could not. Skipping one does not save time; it moves the discovery to the stage where it costs money.
The three questions
| Stage | Question it answers | What it cannot tell you |
|---|---|---|
| Backtest | Did this logic work on past data? | Whether it works now, or whether you can execute it |
| Forward test (paper) | Does it work on data it has never seen? | How you behave when real money moves |
| Live, small | Does the whole system work, including you? | How it performs across a full market cycle |
The third column is the useful one. Each stage has a blind spot, and the next stage exists specifically to cover it.
Stage one: the backtest
A backtest runs your rules over history and reports what would have happened. It is cheap, fast, and the only stage where you can test years in minutes.
Its blind spot is that you chose the rules knowing what the history looked like. That is not dishonesty — it is unavoidable — but it means a good backtest is weaker evidence than it feels.
Stage two: the forward test
A forward test runs the strategy on data it has not seen. Two ways to do it.
- Out-of-sample backtest. Hold back the most recent six months, build on the rest, then test on the held-back period. Fast, and still historical.
- Paper trading. Run it forward in real time on live market data with simulated fills. Slower, and genuinely unseen.
Do both. The out-of-sample test is quick and catches obvious over-fitting. Paper trading catches everything the backtest's assumptions hid — fills, timing, your own configuration mistakes.
What this stage catches
This is where curve fitting gets caught. A strategy tuned to history performs noticeably worse on data it has never seen, and the gap between the two is your over-fitting measurement.
Stage three: live, small
One lot. Real money. For at least a week, preferably two.
What this catches that paper trading cannot:
- Real slippage on your actual order sizes at your actual times.
- Broker behaviour — rejections, margin checks, order type handling.
- The daily session routine, and whether you actually keep up with it.
- Your own reaction to a red day with real money in it. This is the big one.
The cost of skipping each stage
Stated plainly, because skipping is always tempting and the cost is always deferred rather than avoided.
| Skipped | What you find out later | Where |
|---|---|---|
| Backtest | The logic never worked | In live P&L, slowly |
| Out-of-sample | You fitted the parameters to history | In live P&L, confusingly |
| Paper trading | Your configuration was wrong | On day one, expensively |
| Live small | You cannot sit through a drawdown | After scaling up |
The third row is the common one and the most avoidable. A wrong strike rule, an exit time that never fires, a stop set on the wrong leg — these are configuration mistakes, not strategy mistakes, and two weeks of paper trading finds them for nothing.
Minimum time at each stage
| Stage | Minimum | Better |
|---|---|---|
| Backtest | 2 years of data | 3 to 5 years including a volatile period |
| Out-of-sample | 6 months held back | A year |
| Paper trading | 2 weeks | A month, including an expiry cycle |
| Live small | 1 week | 2 to 4 weeks |
These are deliberately modest. The common failure is not testing for too short a period — it is skipping a stage entirely.
When to go back a stage
- Out-of-sample results much worse than in-sample → back to design, not forward
- Paper results far from backtest → find out why before risking capital
- Live results far from paper → usually slippage or configuration, and worth diagnosing
- You changed the strategy at any stage → the earlier evidence no longer applies
The last line is the one people resist. Adjusting parameters after seeing paper results makes the paper test an in-sample test, and you need fresh unseen data again.
The short version
- Backtest asks whether the logic worked on history
- Forward test asks whether it works on data it has not seen
- Live-small asks whether the whole system works, including you
- Each stage covers the previous stage's blind spot
- Changing the strategy mid-way invalidates the evidence you already collected
Frequently asked questions
A backtest runs your rules over history you have already seen. A forward test runs them on data the strategy has not seen, either by holding back recent months or by paper trading in real time.
Because a backtest cannot tell you whether the rules were fitted to the history you looked at, nor whether you can actually execute them. Those are the things the next two stages check.
Two weeks minimum, a month is better because it covers a full expiry cycle. The point is real-time data the strategy has not seen.
Paper trading cannot show you real slippage, real broker behaviour, or how you react to a losing day with real money. The last one matters most.
Your paper test becomes an in-sample test, and the evidence no longer counts as unseen. You need a fresh forward test.
Two years minimum, three to five is better, and it should include at least one genuinely volatile period.
Start with a free 3-day trial
Build a strategy, backtest it and run it on paper — no broker, no IP and no money needed to try it.
