The concepts behind every verdict

A backtest replays a strategy's exact rules over historical price data as if you had traded them, trade by trade, with no hindsight. If the rules say "buy when X, exit when Y", a computer can check every X and Y that ever happened — tens of thousands of situations instead of the five hand-picked examples a video shows you.

What a backtest is NOT: proof of the future. Markets change. A good backtest tells you whether the idea EVER worked and whether its profits survive costs — that's the bar most strategies fail before the future even gets a vote.

Ours run on up to 17 years of 1-minute data with every stop and target checked minute by minute, and when both the stop and target are touched in the same minute, we count the loss. Pessimism is a feature.

"90% win rate" sounds unbeatable. But if the 9 wins make $10 each and the 1 loss costs $100, you're losing money while winning almost every trade. Win rate says nothing without the SIZE of wins versus losses.

The number that actually matters is expectancy: (win% × average win) − (loss% × average loss) − costs. Positive expectancy = edge. Everything else is atmosphere.

This is the most common trick in strategy videos, usually built with a far target and a tight... no, the reverse: a tiny profit target and a wide stop. The account bleeds slowly upward, then one bad week erases it. Our claim-vs-reality scoreboard shows both numbers side by side for exactly this reason.

Every trade pays a toll: the spread (buy price vs sell price), commission, slippage (your order filling at a worse price than you clicked), and swap (financing charges for holding overnight).

For a day-trade on an index, the toll is roughly 2-4 basis points per round trip. That sounds tiny — but scalping strategies often have a TRUE edge smaller than that. They pick up real pennies in front of a cost steamroller.

It's the single most common cause of death in our library: positive gross, negative net. The cost-sensitivity strip on every verdict shows what happens at 0x to 2x realistic costs — an edge that only exists below real costs isn't an edge.

Flip a fair coin 10 times and you might get 7 heads — luck. Flip it 10,000 times and get 7,000 heads, and something real is going on. The t-stat measures exactly that: how far a result is from "just luck", given how many trades and how wild their swings are.

Rule of thumb: |t| below 2 = could easily be luck. Around 2-3 = interesting. Our labs demand roughly 3+ after corrections, because we test many strategies — and testing many things makes lucky ones appear (see overfitting).

A p-value is the same idea flipped: the probability that pure chance produces a result at least this good. p=0.30 means "a coin does this 30% of the time" — not impressive, no matter how good the equity curve looks.

Suppose a strategy made money in a backtest. Was it the ENTRIES, or would anything have made money in a rising market? To find out, we build hundreds of "random twins": same market, same trading sessions, same holding times, same costs — but the entry timing is random.

If the real strategy beats almost all of its twins, the entries carry information. If it lands in the middle of the pack (null-p near 0.5 or higher), the rules are decoration — the profits came from the market drifting, not from skill.

This single test kills more "profitable" strategies than any other check we run. It's also almost never shown in YouTube backtests, for reasons you can now guess.

Take any strategy. Try 50 settings combinations: periods, sessions, stops, filters. Pick the best one. It will look GREAT — and it means almost nothing, because the best of 50 random tries beats the average by pure arithmetic.

That's overfitting (data mining): tuning rules until they memorize the past instead of learning something about markets. The past doesn't repeat its noise, so the memorized version dies on new data. It's how most sold "optimized" strategies are born.

Defenses: decide the rules BEFORE looking, test them on data the optimization never touched (out-of-sample), and apply a mental discount proportional to how many things were tried. Our playground counts your tries and reminds you — because we'd rather teach the disease than sell the symptom.

If you tune a strategy on 2015-2022 data and then test it on 2023-2026 data it has never seen, the second result is the honest one. The data you tuned on is "in-sample"; the untouched data is "out-of-sample".

Every verdict here includes a chronological split: does the most recent 30% of history agree with the first 70%? A strategy that only worked years ago gets called out — profitable-in-2016 doesn't pay 2026 bills.

Serious research goes further: a final "holdout" slice of data stays sealed until the very end and is used exactly once. One look, one answer, no do-overs.

Prop firms sell evaluations: pay a fee ($300-1,100), grow a demo account 8-10% without breaching loss limits, get a "funded" account and a profit split. The pitch: trade their capital. The math: most revenue comes from failed and repeated evaluation fees.

Here's the uncomfortable part our simulations make visible: even a LOSING strategy passes evaluations quite often — short-term luck reaches the target before the negative edge catches up. Passing proves much less than it feels like it proves.

And "funded" is not "paid": the same loss limits keep running inside the funded account, and a meaningful share of funded accounts die before the first payout ever arrives. Our panels simulate the whole journey — evaluation, funded stage, first payout — using the strategy's own trades. The gap between "chance of funding" and "chance of a real payout" is the number the ads never mention.