A tuned EMA cross at +157.8% in training: what the period it never saw says
The typical case: two evenings turning the two settings of a moving-average cross on TradingView until the backtest looks great, and the question of whether to spend a third. Below is the report the paid audit produced on that case, exactly as a buyer reads it.
The starting point
- A Pine v5 script with two settings: fast EMA from 3 to 50, slow EMA from 20 to 200, fees 0.08% per fill.
- Candles: BTCUSDT 4h, 6,570 candles from 18 October 2023 to 17 October 2026, sent as a CSV file, as a buyer does.
- Engine: the paid audit program of the server (Rust, CPU), identical bytes, and the server code that normalizes its report.
Show the script
//@version=5
strategy("EMA Cross Tuned", overlay=true, commission_type=strategy.commission.percent, commission_value=0.08)
fastLength = input.int(12, "Fast EMA", minval=3, maxval=50)
slowLength = input.int(40, "Slow EMA", minval=20, maxval=200)
fast = ta.ema(close, fastLength)
slow = ta.ema(close, slowLength)
if ta.crossover(fast, slow)
strategy.entry("Long", strategy.long)
if ta.crossunder(fast, slow)
strategy.close("Long")What the audit did
- Free check before payment: the script is read; about 9 trades are expected on the set-aside part, where the audit asks for 32. The page announces a likely “insufficient data” verdict before any payment.
- Your tuning, done cleanly: 100 settings of the two inputs (out of 8,192 possible) tried on the oldest 80% of the candles only.
- A single opening of the last 20% (1,314 candles, 12 March to 17 October 2026) with the setting chosen, compared with holding BTC and with an SMA 200 trend filter.
The report, as the buyer reads it
Loading the report…
Formal verdict of the report: INCONCLUSIVE. The private report also lists the fingerprints of the script, the candles, the training part, the set-aside part and the engine, and offers a short PDF receipt in English: the PDF of this example.
For comparison: the original version, without tuning
| Training | Set-aside period | Audit verdict | |
|---|---|---|---|
| Original 12 / 40 | +32.9%, 117 trades | +12.3%, 29 trades (holding: +15.9%) | no evidence of an edge (below holding) |
| Tuned 34 / 200 | +157.8%, 36 trades | +26.0%, 7 trades (holding: +15.9%) | no evidence yet (7 trades for 32) |
Tuning multiplied the training result by about 5 and divided the number of trades by 3. The version that can be judged does less than holding BTC; the tuned version trades too rarely to be judged. Before a third evening on the same two settings, this is what is worth knowing.
Also measured with the same engine: the same script on BTCUSDT 1h (2 years): best training setting −3.55%, set-aside period −11.85%, no evidence of an edge. A finer timeframe does not change the answer here.
Free judge or €9.90 audit?
| Free online judge | Audit, €9.90 once | |
|---|---|---|
| What is judged | your script as written | your tuning: up to 100 settings of your inputs, on training only |
| Candles | BTCUSDT, or your CSV | your CSV (regular candles, no gaps) |
| Compared with | holding the asset, and a test of chance | holding the asset, SMA 200 trend filter, 30% drop limit, 1 trade per 40 candles |
| Kept | in your browser | private report on your account, signed, with PDF; 10,000 Monte Carlo redraws from 10 trades |
The two do not use the same rule: the judge has a test of chance that the audit does not have; the audit guards against chance with its minimum number of trades. Paying never changes a verdict.
Check it yourself
Candles: the last 6,570 rows of the online judge’s public file /static/juge/data/v2/btcusdt-4h.csv (Binance BTCUSDT data, CC BY-NC-SA 4.0 licence), saved as a CSV with a time,open,high,low,close,volume header.
CSV SHA-256 eeb8e17ab78f5a93f4188de8ba7741ad0aaaf7a8a97374c1c86609b09e211017
Pine SHA-256 4110ce3f9e519649919adc3c4bbe9978872a4630b3539e02135842a799191ab0
Engine SHA-256 b928851b7bbccab41fd034d26e6f223ec1c2c13d5a547f9905dd438b5fb894a8
Report fingerprint ebf14b57779415d4c5d7ddfd7f6d06decf0f1abc2383e6683ea30f95a3186f1c
Research on past data, never advice. One period is not a forecast; no result here promises a gain.