Before You Automate a Single Trade
What a backtest can actually tell you
A backtest tells you whether a rule, applied mechanically to the past, would have produced a certain sequence of trades. The rule is internally consistent, it can be coded without ambiguity, and across one slice of history it did what you designed it to do. That is useful. It is also a smaller claim than most people treat it as.
The trouble starts when a backtest gets read as a preview of the future instead of a description of the past. Markets do not repeat themselves the tidy way a backtest assumes: volatility shifts, liquidity changes, and the participants on the other side of your trades are not the ones active five years ago. A backtest can only replay what already happened once.
I have found the most useful way to treat a backtest is as a filter, not a forecast: it knocks out ideas that are obviously broken, rules that do not survive their own logic, parameters that only worked because of one fortunate stretch. What survives still has to earn trust through slower methods.
Overfitting versus a genuinely robust rule
Overfitting happens when a strategy has been tuned until it fits the noise in historical data, not just the signal underneath it. Adjust enough parameters against the same stretch of history and eventually you land on a combination that looks flawless on that data. The strategy did not get smarter. It got closer to memorizing an answer key that will not be handed out again.
The tell is usually complexity that cannot be explained in one plain sentence. If you cannot say why a rule should work, independent of the chart that produced it, stop there. A genuinely robust rule is almost boring to describe: it responds to a condition that makes economic sense, and behaves reasonably across several kinds of markets, not only the one that produced the best curve.
We have noticed that traders who make it through the automation transition tend to distrust their own best backtest result, because the best-looking result in a batch of variations is more likely to be the one that got lucky than the one that is actually sound.
Paper trading is a mindset, not a formality
A lot of people treat paper trading as a box to check before the real work begins: run it for a couple of weeks, confirm the orders fire, move on. That misses what it is actually for.
Paper trading, done properly, is where you find out whether you can tolerate watching a strategy operate. A rule that looked calm on a backtest can produce a string of small losses that feels different in real time than on a chart. It is also where execution details a backtest smooths over, slippage, partial fills, the gap between signal and order, start to show up as real friction.
The mindset shift is treating the forward test as a genuine test, not a formality on the way to going live. If a rule cannot hold up across calm stretches, choppy stretches, and a news-driven move while still only simulated, that is information worth acting on, not an inconvenience to route around.
What small live size actually tests that paper trading cannot
Paper trading removes one variable that matters more than most people expect: consequence. There is a real difference between watching a simulated position move against you and watching your own capital do the same thing. That difference shows up in decisions made under pressure: exactly what you are trying to validate before handing control to a machine.
Small live size is where you test your own behavior around the system, not only the system itself: whether you override it, second-guess a signal you would have taken without hesitation in simulation, or check on it compulsively in a way that would be unsustainable at real size. These are not weaknesses. They are exactly the kind of information a backtest can never surface, because it puts nothing real at stake.
It also tests the plumbing under real conditions: order routing, fills at actual prices, platform behavior during a fast market. None of that is guaranteed to match what a simulation predicted, and small size is what lets you find the gaps while being wrong still costs little.
Position sizing and the failure modes of full automation
Automating a strategy does not remove risk. It changes what the risk looks like, and how fast a problem compounds. Three failure modes are worth sitting with before flipping anything to fully unattended operation.
The first is a bug executing at scale. A manual trader who mis-clicks an order usually notices within seconds. An automated system with a logic error can repeat that mistake hundreds of times before anyone looks at a screen, because the point of automation is that nobody needs to be watching. Position limits, frequency caps, and circuit breakers on abnormal behavior are the difference between a bad hour and a ruined account.
The second is a stale or interrupted data feed. A strategy making decisions off a feed that has silently frozen, or is running seconds behind the market, can look normal while reacting to stale information. Automated systems do not know real data from broken data unless you have built checks for it.
The third is a regime change the strategy has never encountered. A rule built and tested during one kind of market, trending, range-bound, quiet, can behave in ways nobody anticipated once conditions shift. This is exactly what paper trading and small live size give early warning about: a strategy struggling with unfamiliar conditions at a size you can absorb without lasting damage.
A pre-automation checklist, in the form of questions
Before flipping a strategy from backtest to unattended automation, it helps to run through a short set of questions rather than a checklist of tasks. Tasks get completed and forgotten. Questions have to be answered honestly every time you consider raising the size or removing a safeguard.
Can the strategy's logic be explained in plain language to someone who is not a trader, without leaning on the chart that produced it? Has it been watched through more than one kind of market condition? Is it clear what happens if the data feed goes stale or the platform disconnects, and has the system actually been built to handle that? Has it been tested at a size small enough that being wrong would sting but not derail you, with pause conditions written down somewhere other than memory?
None of these questions have a universal right answer. What they share is that they force a distinction between a strategy that performed well in a controlled setting and one you trust enough to run unwatched.
| What it actually tests | What it can't tell you | Discipline it requires | |
|---|---|---|---|
| Backtesting | Whether a rule, applied mechanically to historical data, produces a coherent, executable sequence of trades. | How the strategy behaves under conditions the sample excluded, or how you behave once real consequences are attached. | Resisting the urge to keep adjusting parameters until the result looks better, rather than treating it as a reason to stop. |
| Paper trading (forward test) | Whether the strategy holds up in real time, and whether execution mechanics behave as expected. | How you will react once your own money, not a simulation, is on the line. | Treating a simulated loss as seriously as a real one, and running the test across more than one kind of market. |
| Small live sizing | Your own behavior under real consequence, and whether the platform performs as expected with real capital moving through it. | Whether the strategy keeps working at a size large enough to matter, or once conditions genuinely shift. | Keeping size small enough that a full loss is tolerable, resisting the pull to scale up the moment results look encouraging. |
Is automated trading safer than manual trading?
No. It changes what can go wrong rather than removing risk. A human trading manually can freeze or panic, but also tends to notice quickly when something looks wrong. Automation removes hesitation and most of its benefits at once. It can execute a flawed decision with perfect consistency, which is what makes a bad assumption more dangerous once a system runs unattended.
How long should I paper trade before going live?
There is no fixed number of weeks that makes a strategy ready. What matters more is whether it has been watched across genuinely different conditions, calm, volatile, trending, range-bound, rather than simply letting a calendar pass. A strategy observed only during one uneventful stretch has not really been tested, no matter how many days went by.
Do I need to know how to code to automate a trade?
Not necessarily. Many platforms let you define rules without writing custom code, though you should still understand how those rules translate into actual orders. The bigger risk is not the coding gap, it is not understanding your own logic well enough to know when the system is behaving correctly. Whether you build it yourself or configure a platform, you are responsible for understanding what it does when something goes wrong.
We keep a short set of risk notes for people working through this shift, from a backtest they like to a system they would trust running unattended. If that is useful, get the risk notes below.