We backtested 31,407 crypto trading setups. After costs, nearly two in three lost money.

Sixteen classic indicators on 1-minute to 4-hour bars, on Upbit's Korean-won market, with the exchange's 0.05% fee and one price tick of slippage charged on every trade. A setup is one indicator rule with its settings, sometimes with a filter or an exit rule added. This site maps where the money went.

Share of setups that made money after costs, by bar length

34%

55%

74%

81%

10m

15m

30m

The percentage above each column is the share of setups above water. Vertical scale: yearly return in %.

Each column is one bar length. Its outline shows how the yearly returns of every setup tested on that bar length are spread, from −100% at the bottom to +80% at the top; width grows with the number of setups. Costs: Upbit's 0.05% fee plus one price tick of slippage on every trade. 1- and 3-minute bars: 8 coins, April to October 2026. 5- and 10-minute bars: 12 coins, January 2025 to October 2026. 15 minutes to 4 hours: 12 coins, January 2024 to October 2026.

Return per year after costs, %

Sea level (0) is break-even after costs: blue is under water, green to brown is land.

What the map shows

Short bars drown

On bars of 10 minutes or less, no more than 6% of setups made money after costs. The fee and the tick are the same on every trade, but the moves get smaller as the bars get shorter.

Results by timeframe

Four-hour bars are mostly land

81% of 4-hour setups were above water and 1,267 passed our two-period test. Holding the same coins made +10.6% a year but was down 74% at its worst; most 4-hour setups never fell more than 40%.

Indicator by indicator

Winning often is not making money

Rules that buy oversold dips won more than half of their 4-hour trades, and most of them still lost money. Breakout rules won about a third of their trades and made money, because their wins were several times larger than their losses.

Why win rate misleads

Backtests flatter

The best rule we ran live made +92% a year in replay. On a real server, with paper money, it followed every rule for two weeks and lost 3.93%. Read these numbers as a map of where not to dig, not as a promise.

The two-week server test

The atlas: every indicator on every bar length

Median yearly return after costs of each indicator's single-indicator settings. Pick an indicator to see every setting, and what adding filters and exit rules did.

Indicator	1m	3m	5m	10m	15m	30m	1h	4h
Moving-average cross	−100	−100	−100	−94	−72	−22	+15	+37
Price above an average	−100	−100	−100	−100	−100	−96	−70	+25
RSI, buying strength	−100	−100	−100	−96	−83	−51	−12	+40
RSI, buying oversold	−96	−47	−60	−42	−16	−14	−14	−6
MACD	−100	−100	−100	−100	−100	−96	−74	+8
Bollinger breakout	−100	−100	−100	−97	−82	−47	0	+53
Bollinger, buying the lower band	−100	−98	−91	−72	−44	−30	−26	−21
Donchian breakout	−100	−98	−96	−79	−61	−20	0	+42
Keltner breakout	−100	−100	−100	−97	−84	−50	−8	+50
SuperTrend	−100	−100	−100	−98	−88	−60	−17	+34
CCI	−100	−100	−100	−95	−79	−48	−21	+13
Rate of change	−100	−100	−100	−100	−100	−97	−75	+18
ADX	−100	−100	−100	−100	−100	−91	−63	+7
Stochastic	−100	−99	−97	−86	−69	−51	−41	−30
Williams %R	−100	−100	−100	−99	−87	−62	−43	−36
VWAP	−100	−100	−100	−98	−93	−85	−74	−60

Research

Six weeks of building, running and auditing trading systems with an AI assistant, written up with every loss included.

Six weeks of letting AI build a crypto trading system
−44% in the first week, ₩25.2 million of paper losses across 918 exits, and a rule-perfect server that still lost.
The best backtest we found, and why we did not trust it
A 4-hour Bollinger breakout made +63% a year after costs in replay. Here is everything that could still be wrong with it.
The one-minute backtest trap
Fees-only backtests on 1-minute bars showed returns in the tens of thousands of percent. One price tick made them all disappear.
Can a language model pick crypto trades?
6,536 trade candidates, 87 independent events: the trades the model liked most did worst.

All research

Tools

Can a trading bot beat fees?
Edge per trade against what each trade costs.
Break-even win rate
How often you must win, given your average win, loss and costs.
Trades needed to prove a strategy
How many trades it takes before a result means anything.
Upbit tick size and cost
What one price step costs at any price on Upbit's won market.
Crypto profit calculator
Profit after fees and the break-even sell price.
Position size calculator
How much to buy for a fixed risk per trade.

All tools

About the data

Prices come from Upbit's public candle data. Every order fills at the next bar's open. Each setup trades the same 12 large coins (8 on 1- and 3-minute bars) with equal money per coin. Settings were judged on a first period and checked on a second period they had never seen. 2,098 setups whose filters never allowed a trade are left out of the shares. The coins were chosen in 2026 among those listed for at least two years, which flatters every result here; the methodology lists this and the other limits.

All 31,407 results are free to download under CC BY 4.0 from the data page, or to browse in the explorer.