Short definitions of the terms you meet in a backtest, at your broker and in prop firm rules. Each entry carries one number from our tests, so you can see how the term behaves on real data.
The numbers come from the published chapters of the Strategy graveyard and describe our reconstructions on our data, after costs.
1R is the amount you lose if price reaches your stop. A trade’s result divided by 1R is its result in R, and the average of a series estimates expectancy. With this one unit you can compare every trade and every strategy, whatever the size of your account or position.
Read →A trailing drawdown is a maximum loss limit that trails your account: a floor that rises behind the highest level reached so far and never comes back down. The end-of-day (EOD) version is milder: it counts only the balance after the daily close, while the intraday version counts every peak. In our simulation, the same trades with no edge passed the phase 51 times in 100 with a static limit and 41 times with an EOD trailing limit.
Read →A daily loss limit counts from your account at the daily reset, usually from the balance or the higher of balance and equity; equity triggers the breach. Usually the reset is at midnight server time or at 17:00 in New York. That is 23:00 in Central Europe and 22:00 in the weeks when US and European clock changes differ. In our simulation, starting capital versus balance as the base mattered only with a limit of exactly three stops. Pausing for the rest of the day instead of losing the account raised the pass rate from 32% to 54%.
Read →Gambling is a game with a negative expected value, where you have no influence over the odds. Trading without an edge after costs is mathematically similar. Before costs it averages zero; after costs it loses, on average, what a trade costs. Trading with an edge after costs is not gambling in this sense. So the question is not “whether”, but “does your series differ from chance?” You can check that with numbers.
Read →A trade’s result in multiples of the risk: −1 R is a loss equal to the distance to the stop, +2 R twice that as profit. It lets you compare trades with different stops; a loss beyond −1 R means a gap or slippage.
In our tests
In our channel-breakout test, 7,523 trades on gold and eight indices added up to +149 R, even though each instrument has a different price and stop (chapter 1).
The average result of a single trade, ideally in R and after costs. A positive average is only the start: you still have to check that it is not chance.
In our tests
Across 1,344 variants and over a million trades in 2018–2022, the average trade came to −0.060 R, because costs take their toll (chapter 7).
A positive expectancy after all costs that repeats outside the sample where it was found. A positive series is only a candidate: what separates it from luck is the number of trades, the result without the best trades and a comparison with random entries.
In our tests
The 13 best of 1,344 rule variants made +0.201 R on average in the training years and +0.049 R in the years after (chapter 7).
Total gains divided by total losses: a PF of 1.00 is break-even, below 1 the strategy loses. A backtest PF is easy to overrate, because the best of many variants looks good even without an edge.
In our tests
If none of 224 variants had an edge, the best one would still show a PF of about 1.3 over about 300 trades (chapter 7).
The share of trades that close in profit. On its own it says little: with a 1:2 risk-reward ratio, 33.3% is enough to break even before costs.
In our tests
In our ORB test on gold, the variant with 49.7% profitable trades averaged −0.013 R, while the one with 37.0% made +0.034 R (926 sessions, chapter 10).
The distance to the target divided by the distance to the stop: 1:2 means the target is twice as far as the stop. The win rate you need to break even before costs is 1 ÷ (1 + RR).
In our tests
In two signal channels, targets were spaced roughly $5 apart, or 0.3–0.5 R, with a stop of $10–13 (chapter 45).
Several losing trades in a row. With a low win rate, long streaks are normal; before changing your rules, work out how often they happen by chance alone.
In our tests
In a channel with 89 signals and a 45% hit rate, the longest losing streak was 15 signals (chapter 45).
The fall in capital from a peak to the lowest point before a new high, in percent or in R. History shows only one order of trades; the same trades in random order often produce a deeper drawdown.
In our tests
In our carry test on 28 currency pairs, the drawdown reached 37.6% and lasted from July 2007 to January 2009 (chapter 38).
The average result divided by its random error; in a single test, t ≈ 2 is the usual threshold of significance. If you test many variants, the best one passes 2 even with no edge at all.
In our tests
With 224 variants per instrument, the best t-statistic ranged from 1.99 to 3.01, roughly what chance alone predicted (chapter 7).
The share of random draws that give a result at least as good as the signal. The closer to zero, the harder it is to explain the result by chance; p = 0.20 means chance matches it one time in five.
In our tests
In a channel with 89 signals, a random direction did at least as well in 20% of draws, so p = 0.20 (chapter 45).
The number of trades at which your average can be told apart from zero. Roughly n ≈ (2 × SD ÷ average)², where SD is the standard deviation of results in R; a small edge needs hundreds of trades.
Worked example
An average of +0.1 R with a standard deviation of 1.5 R gives (2 × 1.5 ÷ 0.1)² = 900 trades; an average of +0.2 R needs only about 225.
The same result after removing the 3 or 5 best trades. If it then changes sign, the strategy hangs on a few lucky trades that need not repeat.
In our tests
A cell chosen after seeing the result made a PF of 1.15 on 49 trades in the unseen year 2026, but only 0.33 without the three best trades (chapter 16).
The average annual result divided by its volatility; 0 means no edge. It does not show whether losses come rarely and all at once, so read it together with drawdown.
In our tests
The carry strategy on 28 pairs had a Sharpe of 0.30 in 2005–2026, against our threshold of 0.40 (1,133 weeks, chapter 38).
The average annual profit divided by the largest drawdown. It tells you how many years of profit the worst period costs: a Calmar of 0.33 is about three years of recovery.
In our tests
In our carry test, a Calmar of 0.08 meant almost 13 years of recovery: 2.9% profit a year against a 37.6% drawdown (chapter 38).
The difference between the buy price (ask) and the sell price (bid), paid on every trade. It is easy to understate in historical data, so measure it on ticks or on your own trades.
In our tests
On gold, the median spread from 3.57 million ticks was $0.09, while the spread column stored with the candles showed $0.05, or 56% of the real value (chapter 10).
The broker’s extra charge above the market price, hidden in entry and exit prices and, for overnight positions, in the swap. Tick history does not include it, so check the real cost on your own trades.
In our tests
On gold, the cost measured from ticks was $0.09, while the real cost, including the account markup, was about $0.19 (chapter 26).
A fee or credit for holding a position overnight, coming from the interest-rate difference or the cost of funding. The broker adds its markup, and once a week (often on Wednesday) charges a triple swap for the weekend.
In our tests
Judging by today’s swaps on 28 currency pairs, the broker keeps 0.59 to 2.14% a year per side, median 1.01% (chapter 38).
The difference between the price you expect and the price your order actually fills at. It grows in fast moves and around major data, so count costs from those moments separately.
In our tests
In our ORB test on gold, the round-trip cost of $0.13 per trade was a median spread of $0.09 plus slippage (chapter 10).
A trade’s cost divided by the distance to the stop, i.e. the cost in R. With a cost of 8% of R, a strategy has to earn more than 0.08 R per trade before costs just to break even.
In our tests
On bitcoin, the spread alone ate one fifth to one quarter of the risk, so even random entries averaged about −0.25 R per trade (chapter 5).
The average range of a candle, usually over the last 14, a measure of volatility in price points. Stops, targets and costs expressed in ATR can be compared across instruments and years.
In our tests
In our level tests, the stop sat 1 ATR beyond the level and levels counted only up to 2 ATR from price, so the same rule worked on gold, indices and USD/JPY (chapter 14).
Size the position from the risk: lot = risk amount ÷ loss of one lot at the stop. Then every 1 R is the same percentage of the account, and a drawdown in R converts easily to percent.
In our tests
In our ORB test, a simulated drawdown of 40–64 R meant losing 20–32% of capital at a risk of 0.5% per trade (chapter 10).
The most a prop firm lets you lose in one day, usually a percentage of the account. The details matter: what it is measured from (starting capital, balance or equity at the start of the day) and when the day resets.
Worked example
A 100,000 account with a 5% daily limit and a start-of-day balance of 110,000. Measured from that balance, today’s floor is 104,500; measured from the starting capital, 105,000.
The lowest account level below which the prop firm ends the challenge or the account, usually a percentage of starting capital. It can be static or trail the highest balance.
In our tests
One bad carry week from 2007, −9.3% at 0.5% risk per position, would have eaten almost the entire typical drawdown limit of a challenge (chapter 38).
A maximum loss limit that moves up behind the highest balance (at end of day or intraday) instead of staying fixed. The same result can therefore stay within a static limit and breach a trailing one.
Worked example
A 100,000 account with a 6,000 limit. When the end-of-day balance reaches 104,000, a trailing limit moves from 94,000 up to 98,000, while a static limit stays at 94,000.
The moment from which a prop firm counts a new day for the daily limit, usually midnight server time or a fixed hour in the firm’s time zone. Your local reset time can shift when the clocks change.
In our tests
For a trader in Poland, the gap between server time and local time is two hours instead of one for 3–4 weeks a year (chapter 2).
Balance counts only closed trades; equity adds the running result of open positions. A limit based on equity counts a floating loss before you close the trade.
Worked example
A balance of 100,000 and an open position at −1,500 give equity of 98,500; a limit based on equity counts that loss before you close the trade.
A prop firm rule that your best day may account for at most a set share of total profit. One very good day can therefore make a challenge longer instead of shorter.
Worked example
A 10,000 profit and a 40% rule. The best day may contribute at most 4,000, so after a +5,000 day the total profit has to grow to 12,500.
A backtest error in which a decision uses information that did not exist yet at that moment, such as the close of a daily candle still in progress. It inflates the result, so few people look for it.
In our tests
On the daily chart, the backtest used a candle other than the last closed one in 98.3% of decisions (17,783 of 18,085), lifting PF from 1.03 to 1.29 (chapter 1).
An indicator changes or removes signals after the fact, so the chart history looks better than what you could see live. Judge it by what is visible at the moment of decision, not in the evening.
In our tests
The same arrow signals gave a PF of 1.33 on the saved history and 0.54 on real ticks (1,293 entries, chapter 5).
Overestimating the best of many tested variants: when you pick the best, you mostly pick the luckiest. Every filter added after the fact is one more variant in that search.
In our tests
The 13 best of 1,344 variants averaged +0.201 R in training and +0.049 R in validation, about three quarters less (chapter 7).
Split the data in advance: search for the rule in the training sample, and open the holdout sample, set aside from the start, once and without changes. Adjusting the rule afterwards turns it into more training data.
In our tests
To tell a Sharpe from zero at p < 0.05, about 0.36 is enough with 21 years of data, while a 7.7-year holdout sample would need 0.59 (chapter 38).
The same procedure at moments when the cause under study is absent, such as random entries with the same stop and target. If the placebo does about as well, geometry and cost drive the result, not the signal.
In our tests
The arrows made −0.298 R per trade and random entries at the same hours −0.246 R, so the signal was worse than chance (1,293 entries, chapter 5).
A test that randomly shuffles or flips labels many times (for example, the direction of whole days) and counts how often chance does at least as well. Drawing in blocks keeps the dependence within a day.
In our tests
For the arrows, a permutation in daily blocks (2,000 draws) gave p = 1.00: randomly flipping the direction of whole days did at least as well (chapter 5).
A correction for the fact that with many tests at once, some will come out “significant” by chance. The Benjamini–Hochberg procedure lowers the p threshold to control the share of false discoveries (FDR).
In our tests
The best of 12 cells had p = 0.079 before the correction and p = 0.94 after correcting for 12 tests (chapter 16).
A look-ahead check: cut the data exactly at the moment of decision, recompute the indicators and compare them with the values from the full data. Any difference means the full data are giving something away.
In our tests
In our automatic truncation test, 227 signals × 44 features came out identical on full and truncated data (chapter 1).
The time next to candles is the broker’s server time, stored as if it were UTC. At our broker it is New York + 7 hours, with clock changes following the US.
In our tests
When we took server time for UTC, the New York session gave 3 trades in six years and the placebo 127 (chapter 2).
The simple alternative a strategy has to beat at the same risk: buying at the same moments and holding. In a rising market, many rules “work” only because they buy.
In our tests
The best bitcoin rule made +248 R on 1,008 trades in 2023–2025, less than the +261 R from “always buy” at the same moments (chapter 7).
The distortion you get when you see only what survived: arrows left on the chart, signals still listed, messages not deleted. The vanished failures flatter a picture you cannot repeat.
In our tests
In the channel with 37 signals, more than 90% of message numbers were missing, so the visible history was the author’s selection (chapter 45).
An order that closes the position at a loss when price reaches a set level. The distance to the stop defines 1 R, so without a stop there is no result in R.
In our tests
In the first channel, 800 messages had a direction and a price but no stop; we skipped them, because a copier would not know where to put it (chapter 45).
An order that closes the position in profit at a set level (the target). Count how often the target is really hit: a target that rarely triggers turns the strategy into a time exit.
In our tests
The 5 R target was hit in 29 of 926 London sessions (3.1%) and in none of 938 in New York (chapter 10).
A pending order that buys below the current price or sells above it and fills when price reaches it. It tends to fill precisely when the market is moving against you.
In our tests
On eight instruments, a limit order at a level in the trend direction did worse than a market entry in 9 comparisons out of 11 (chapter 14).
An order executed immediately at the current price: buys at the ask, sells at the bid. It is the natural benchmark for limit orders placed at the same moments.
In our tests
In our level tests, a market entry at the same moments did better than the limit order in 9 comparisons out of 11, though it had no edge of its own either (chapter 14).
A pending order that buys above the current price (buy stop) or sells below it (sell stop); when price reaches it, it becomes a market order. It is used to enter on breakouts and also works as a stop loss.
Worked example
Price at 2,000 and a buy stop at 2,005. When price reaches 2,005, the order becomes a market buy, and in a fast move it may fill at, say, 2,005.40.
A stop that follows price at a set distance, for example 1 ATR, and moves only in the direction of profit. It turns big losses into smaller ones but also cuts off some of the big moves.
In our tests
In our ORB test, the trailing stop lowered the average in both sessions; in New York to −0.022 R against +0.006 R with the 5 R target (938 sessions, chapter 10).
Moving the stop to the entry price, usually after a 1 R profit. It stops a winner from becoming a loser, but closes some trades at zero just before the move continues.
In our tests
In our ORB test, a stop moved to entry or trailed closed 450 of the 926 London positions (chapter 10).
A jump in price with no quotes in between, for example after a weekend or a central bank decision. The stop then fills at the first available price, so the loss can be several times 1 R.
In our tests
In our carry test, a stop on the daily close did not protect against a gap: the worst single position lost 6.8 R (chapter 38).