Start here

Backtest

Replaying a strategy over past data to see what it would have done.

Feed historical candles through the strategy one at a time and record the trades it would have taken. Done carefully -- with next-bar execution, honest fees, and pessimistic assumptions about which of the stop and the target was hit first -- it is useful evidence.

Watch out: A backtest is the best case, not the expected case. It knows exactly which market it is trading and never suffers an outage, a bad fill, or a moment of doubt.

Confidence interval

The range the true value plausibly sits in, given this much data.

A single number like '+0.15R' hides how uncertain it is. The interval shows the range consistent with the data collected. '+0.15R [-0.40 .. +0.70]' means the real long-run value could plausibly be anywhere in that span -- including well below zero.

If the interval crosses zero, the honest reading is that you cannot yet tell profit from loss.

Watch out: Wide intervals are not a presentation problem to be tidied away. They are the finding: not enough evidence yet.

Maximum drawdown

The worst peak-to-trough fall in account value.

How far the account fell from its highest point before recovering. If it grew to $1,000 then dropped to $740 before rising again, the maximum drawdown was 26%.

This is the number that decides whether you can actually follow a strategy. A 40% drawdown is easy to accept on a chart and very hard to sit through with real money.

Watch out: The worst drawdown you have seen so far is not the worst possible. It is simply the worst that happened to occur in the period measured, and the next one can be larger.

Expectancy

Average profit or loss per trade, measured in R.

The average outcome of every trade, good and bad, expressed in R. Positive means the strategy made money over the trades measured; negative means it lost. It is an average of what already happened, not a forecast of what comes next.

Watch out: An average over 10 trades and an average over 500 trades look identical on screen and mean completely different things. Always read the trade count next to it.

Fees and the cost floor

What you pay to trade. At small size this dominates everything.

Kraken charges up to 0.80% to take liquidity, on both entry and exit. On a round trip that is roughly 1.6% of position value, paid regardless of whether the trade wins.

For a small account with a tight stop, that fee can be a large fraction of the amount being risked -- so the trade starts well behind and the strategy has to be right often enough to climb back out before it makes anything.

Watch out: Backtests that use the wrong fee look far better than reality. This system had exactly that bug: cached values of 0.16%/0.26% against a real 0.40%/0.80%, understating cost threefold.

Kill switch

A master stop that halts all trading immediately.

When engaged, no new position can be opened by anyone. It is checked before every action and fails closed -- if its state cannot be read, the system treats it as engaged and refuses to trade.

Watch out: It stops NEW trades. It does not close positions you already hold; those keep their stops so they exit on their own terms.

Overfitting

Tuning a strategy until it fits the past, not the market.

Try enough settings on the same history and something will look excellent purely by chance. That result describes the past you searched, not the future you will trade.

The defences are holding data back and never touching it until the end, testing whether nearby settings also work (a real edge is a plateau, not a spike), and counting how many things you tried.

Watch out: This is the single most common reason a strategy that backtested beautifully loses money live.

Paper trading

Trading simulated money against real live prices.

Orders are priced against the real market but no money moves. It tests whether the system works -- connections, restarts, stops firing, reconnecting after an outage -- and lets you watch a strategy behave in real time without paying for the lesson.

Watch out: Paper results are not live results. Real orders move the market slightly, fills are worse than simulated ones, and no simulation reproduces how differently people behave when the money is real.

Profit factor

Gross wins divided by gross losses. Above 1 made money.

Add up every winning trade, add up every losing trade, divide the first by the second. 2.0 means you won two dollars for every one lost. Below 1.0 means the opposite.

Watch out: This is the single most misleading number on any trading screen. On a handful of trades one lucky winner produces a spectacular profit factor that tells you nothing. It is also undefined when there are no losses yet -- which reads as infinite success and actually means 'not enough data'.

R (R-multiple)

One R is the amount you decided to risk on a trade.

Before entering, you choose a stop-loss: the price at which you accept you were wrong and get out. The distance between your entry and that stop, times your position size, is 1R -- the money you put at risk. A trade that makes twice what you risked is +2R; one that hits the stop is -1R.

Measuring in R rather than dollars is what lets you compare a trade on a $200 account with one on a $20,000 account, and it is why position size cannot flatter the results.

Watch out: R only means anything if the stop is real. A strategy with no stop has no R, and a strategy that moves its stop after entry is measuring against a number it changed.

The random baseline

What entering at random achieves, with identical stops and fees.

Before believing a strategy has an edge, compare it against picking entry points at random using the same stops, sizing and fees. Anything it does not beat is coming from the exit and risk rules, not from the signal.

Measured on BTC out-of-sample: random entry returns about -0.44R per trade. That is the fee floor, and it is where every strategy starts.

Watch out: Two of the five strategies tested here performed WORSE than random entry. A strategy can be genuinely worse than no strategy.

Number of trades

How many trades the result is based on. Usually too few.

Trading results are noisy, so distinguishing skill from luck takes far more trades than people expect. Detecting a small real edge needs hundreds. Under 30 trades, almost nothing can be concluded in either direction.

This is why a strategy with a brilliant-looking record over 12 trades deserves no confidence at all.

Watch out: 'Not statistically significant' means the experiment could not tell. It does NOT mean the strategy does not work, and it does not mean it does.