Every backtest anyone has ever shown me had a beautiful equity curve, and that fact alone should bother you more than it usually does. Nobody shows off the ugly ones. If the curve looked bad, they adjusted the rules until it looked good, and that adjustment process is exactly what quietly drains the evidence out of the result. So when someone asks whether their strategy is ready to trade, I skip past the backtest and ask how long the same rules have been running on data that did not exist when the rules were written.
Backtesting and forward testing get lumped together as testing, but they answer different questions. A backtest asks whether your rules would have worked on the past. A forward test asks whether they hold up against a future the rules have never seen. You need both, they catch different failure modes, and confusing the two is how people end up funding strategies that were never actually validated.
What a backtest can prove, and what it quietly assumes
A backtest proves less than most people want it to, but what it does prove is useful. First, it proves your strategy is mechanical enough to be simulated at all. That filter alone kills most trading ideas, because 'buy when it looks strong' cannot be backtested, and forcing yourself to write rules a computer could execute without you in the room is where a lot of the real work happens. Second, it gives you the historical shape of the thing, how often it trades, how deep and how long the drawdowns ran, whether the returns came from three lucky trades or three hundred ordinary ones. Third, it lets you subtract costs. A surprising number of strategies have positive expectancy before fees and negative expectancy after fees, funding, and spread, and a backtest with honest cost assumptions is the cheapest way to find that out.
What a backtest cannot prove is that the edge is real. The core problem is that you looked at the data before you wrote the rules. Even if you never ran a formal optimizer, your brain did the curve fitting for you, because every idea you had was already shaped by charts you had seen. Layer on the classic mechanical sins, lookahead bias where an indicator uses information from inside the bar it is trading, survivorship bias where your asset universe only contains the coins and stocks that made it, and fill assumptions where the simulator grants you limit fills a real order book never would, and you get equity curves that are partly signal and mostly flattering fiction. A good backtest is an argument that the idea deserves a forward test, and honestly that is about all it has ever been.
What forward testing catches that history never will
A forward test runs the exact same rules on live market data with no money attached. Paper trading, if you prefer the older name, though the useful version is stricter than what most people mean by paper trading, because the rules are frozen. No tweaks mid-run, no deciding after the fact that you obviously would have skipped that one trade. The moment you edit the rules, the clock restarts at zero.
The main thing a forward test buys you is data your rules could not possibly have been fit to. Whatever the strategy does over the coming months, it does it against price action that did not exist when you wrote the logic, which makes each forward trade a different grade of evidence than a backtested one.
It also surfaces the ugly practical failures fast. Lookahead bugs reveal themselves within days, usually as signals that appear intrabar and then vanish when the candle closes. You find out that your live data feed disagrees with your historical feed in small maddening ways, that your alerts arrive noticeably after the move they describe, and that the price you signal at and the price you could realistically get are further apart on thin pairs than any simulator admitted. And it tests you. Watching a paper drawdown unfold one day at a time, with no ability to scroll ahead to the recovery, feels completely different from glancing past the same drawdown on a historical chart, and plenty of people discover in this stage that they cannot sit through the strategy they built.
It is worth being honest about what forward testing still misses. No money at risk means no real slippage, no partial fills, no exchange outage at the worst possible moment, and none of the fear that shows up once losses are denominated in your own savings. A clean forward test earns the strategy small size, and small size is really a third test, the one where fills and emotions finally enter the data.
How long before the forward test means anything
Count completed trades rather than calendar weeks. A forward test that produced six trades has told you almost nothing no matter how many months it took, while one that produced eighty trades in three weeks is already interesting. As a rough rule, somewhere around thirty completed trades is the floor before the results are worth glancing at, and closer to a hundred before I would treat them as evidence. A strategy that signals a few times a week typically needs several months to get there, which is exactly why people skip this stage. It is boring, and the whole time the backtest sits there whispering that the money is already proven.
Trade count is one axis, and regime coverage is the other. You want the forward window to contain at least one stretch where the strategy struggled. A trend system forward tested entirely inside a smooth uptrend has been tested against one kind of weather, so if the window never included chop, a sharp reversal, or a dead stretch of volume, extend it until it does.
Then compare distributions rather than headlines. Forward results will almost always come in worse than the backtest, and that degradation is normal enough that its absence would make me suspicious. The real question is whether the forward period could plausibly have been drawn from the same strategy the backtest describes. Take rolling windows of the same length from your historical results and see where the forward window ranks among them. If it sits inside the historical range, weak stretches included, the model is holding. If the forward drawdown is roughly twice as deep as anything the backtest ever produced, something in the simulation was wrong, usually costs or fills, and more waiting will not fix it.
The milestone that unlocks real money
Here is the checklist I use before a strategy touches capital. Every line is there because skipping it has burned someone I know, occasionally me.
- The rules are fully mechanical and stayed frozen for the entire forward window, with zero discretionary overrides.
- The backtest still shows positive expectancy after realistic fees, spread, funding, and pessimistic slippage.
- The forward test covers at least thirty trades, ideally far more, and includes at least one losing streak.
- Forward performance falls inside the historical range of the backtest's rolling windows. Somewhat worse is fine, completely outside the range is disqualifying.
- Every signal was logged, including the ones that were tempting to skip, and the log matches what the rules said to do.
Pass all of that and what you unlock is small live size. Real capital is its own validation stage, because genuine fills and genuine fear both degrade performance in ways paper never captures, so the first stretch of live trading is best treated as a continuation of the test, with tuition attached. Scale up only when the live results rhyme with the forward results. The backtesting side of Blockcircle is built around this exact pipeline, frozen rules, honest costs, forward runs on live data, mostly because we got tired of watching pretty curves fall apart on contact with an order book.
Most strategies die somewhere in this process, and in my experience the majority die in the forward stage, which sounds discouraging until you price the alternative. A strategy that fails on paper costs you a few months of patience. The same failure discovered live costs the patience plus the money, so the forward test is the cheapest exit you will ever get from a bad idea.