Forward testing: the gate a backtest cannot be
Paper trading is not a formality before going live. It is the only test that runs against data nobody has, including you.
A backtest and a forward test sound like the same thing done at different times. They are not. They fail differently, and that is exactly why you want both.
A backtest runs against history. However carefully you hold data back, that history already exists, which means it has already been looked at — by you, by everyone writing about the market, and by whoever designed the indicator you are using. The structure you found in it might be real, but you cannot rule out that you found it because it was there to be found.
A forward test runs against data that does not exist yet. Nobody has seen it. Nothing has been fitted to it. It cannot be peeked at, because there is nothing to peek at. That property is unique and it is worth waiting for.
What a forward test is really testing
Most traders think they are testing the strategy. They are, but it is not the most valuable thing happening.
A forward test exercises the entire path from a signal to a recorded fill, in real time, with everything that can go wrong in between actually in place. The strategy is one component of that path, and in practice it is the component least likely to be broken.
What forward testing routinely catches:
- Timing you had not thought about. A rule that needs the previous day's close behaves differently when "the previous day's close" is not available until the data vendor publishes it. In a backtest it was simply a column.
- Instruments that stop existing. Your test traded an expiry. Live, that expiry rolls, and something has to choose the next one. A test with a continuous series has silently answered a question your live setup has to answer explicitly.
- Plumbing faults. Signals that arrive twice. Signals that do not arrive. A position the engine thinks it holds. A risk limit that reads as zero and refuses everything. These are not strategy problems and no backtest can show them — the module on automation is entirely about this class.
- The rate at which you actually get signals. A backtest says 140 trades over two years. A month of forward testing says you have had three, and two of them were on the same day. That is information about whether you can live with the strategy at all.
- Your own behaviour, partially. Watching a simulated position lose money for three weeks is not the same as watching real money do it, but it is not nothing. Many strategies are abandoned during forward testing, which is the cheapest possible place to abandon them.
How long, and judged on what
The honest answer is: until you have enough trades to say something, not until you feel ready.
The arithmetic is unforgiving. If your strategy makes money on 55% of trades, then after twenty trades there is a very real chance you are looking at nine winners and eleven losers through no fault of the strategy. Twenty trades cannot distinguish a 55% edge from a coin. A hundred barely can.
So set the gate in trades, decided in advance, and write the number down before you start. Thirty trades to confirm nothing is structurally broken. A hundred or more before you give the result any weight as evidence about the edge itself.
And judge it against what the backtest predicted for that same period, not against zero. "Did it make money" is a weak question — a month can be good or bad for reasons that have nothing to do with you. "Did it behave the way the test said it would" is the strong one. Same frequency of trades? Same average size of win and loss? Same worst day? A strategy that made money in a way the backtest never described has not been validated; it has got lucky in a new way.
A backtest over two years: 148 trades, 58% of them winners, average win ₹2,100, average loss ₹1,650, worst day −₹5,400.
Forward tested for six weeks: 11 trades, 7 winners, average win ₹2,250, average loss ₹1,580, worst day −₹5,900.
Eleven trades say nothing about the 58%. But every other number is in the right neighbourhood, and the worst day did not exceed what the test described by a margin that suggests the loss model is wrong. That is a pass on the only question six weeks can answer: is this thing behaving like the thing I tested?
Now compare a different outcome. Same six weeks, 11 trades, 8 winners, but the average win is ₹5,800 and the worst day is −₹14,000. This looks better and is far more worrying. The strategy is taking positions the backtest did not describe, which means something in the live path differs from the test — sizing, instrument, or a rule that is not firing as intended. Profitable and unexplained is a reason to stop, not to go live.
Where paper trading tells the truth, and where it cannot
Simulated trading is honest about rules and dishonest about fills, and it is important to know exactly which is which.
Your simulated order is matched against a price from the feed. A real order is matched against the order book, where your size has to find a counterparty, and where between your decision and the exchange's acknowledgement there are tens of milliseconds in which the price can move. Simulation cannot reproduce that. On a liquid index option at small size the difference is modest. On a far strike, in the last minutes before expiry, or at a size that is a meaningful part of the book, it can be larger than your entire edge.
A simulated book is also a book with no broker in it. There is no margin call, no rejection because the exchange froze the quantity, no order that came back with a cryptic code. Those are real parts of trading that a paper account cannot show you, and they are a good part of why the third gate — live at the smallest size you can place — exists at all.
So treat paper trading as a test of your strategy and your plumbing, and treat the first live weeks as the test of your execution. Trying to get both from one gate is how people end up believing a number that was never measured.
What to take from this chapter
- A forward test runs on data nobody has seen. Nothing can be fitted to it, which is a property no backtest can ever have.
- It tests the whole path, not just the strategy — and the strategy is rarely the broken part.
- Set the gate in trades and write the number down before you start. Thirty to confirm nothing is broken; a hundred before the result is evidence.
- Judge it against what the backtest predicted for that period, not against zero.
- Profitable but behaving differently from the test is a reason to stop, not a reason to go live.
- Paper trading is honest about rules and plumbing, and cannot be honest about fills. Your first live weeks at minimum size are the test of execution.