Oct 30, 2026 · 4 min read
We tuned 92 moving-average settings on six months of BTC. 49 beat buy and hold. In the next six months, with the same settings, 4 did. Here are the four steps and the numbers.
TradingWizard AI5 min read
To trade with AI algorithms, write down one set of rules, test it on past prices it was not tuned on, run it with simulated money, and only then risk a small amount with a hard loss limit. The step most people skip is the second one: we tuned 92 moving-average settings on six months of Bitcoin, and the number that beat simply holding fell from 49 to 4 in the next six months.
An AI model can help at every step: writing the rule, coding the test, or reading the chart for you. It cannot skip step 2 for you.
We took one year of hourly BTC-USD candles from Coinbase and split it in two halves. We ran 92 versions of the same rule, an EMA crossover with the fast average between 3 and 50 hours and the slow one between 20 and 200 hours. Long only, no leverage, 0.1% fee on every buy and every sell.
| First half (tuning) | Second half (test) | |
|---|---|---|
| Dates | 2 Nov 2025 to 29 Apr 2026 | 29 Apr to 24 Oct 2026 |
| BTC price | $110,562 to $75,898 | $75,770 to $80,770 |
| Buy and hold | −31.4% | +6.6% |
| Settings that made money | 0 of 92 | 31 of 92 |
| Settings that beat buy and hold | 49 of 92 | 4 of 92 |
| Median setting | −30.5% | −2.3% |
In the first half every setting lost money, but more than half of them lost less than holding. That looks like a rule that protects you in a falling market. In the second half the market rose a little, and 88 of the 92 settings did worse than doing nothing.
The best setting from the tuning half was EMA 25/40:
| EMA 25/40 | First half | Second half |
|---|---|---|
| Result after fees | −17.1% | +3.0% |
| Round trips | 39 | 39 |
| Buy and hold | −31.4% | +6.6% |
| Rank among 92 settings | 1st | 10th |
If you had stopped after the first half, you would have picked a rule that "beat the market by 14 points". Run unchanged on the next six months, it trailed holding by 3.6 points. Nothing about the rule changed. The market did.
TradingWizard runs step 3 for you. Pick a market and an asset, set up a trading bot, and watch what it does. The bots use Wiz, Elliott Wave, Fibonacci, Cycle, MACD, Moving Average Cross and Gann swings. Each bot shows what it read, why it traded or waited, and when it checks next. Demo trades use simulated money. No real orders are sent.
Put a Moving Average Cross bot on BTC in /wiz and note the date of its first decision, so you know where your own test half begins.
Coinbase BTC-USD hourly candles from the public API, 25 October 2025 to 24 October 2026: 8,750 candles, 10 missing from the feed. The first 200 hours warm up the averages. Each half has 4,275 hours. Trades fill at the hourly close, long only, spot, no stop-loss and no slippage. A position still open at the end of a half is closed at its last close. The 92 settings are every pair from fast 3, 5, 8, 10, 12, 15, 20, 25, 30, 40, 50 and slow 20, 30, 40, 50, 60, 80, 100, 120, 150, 200 where the slow average is more than 1.5 times the fast one. These are past results for one coin over one year and say nothing certain about the next one. This page explains a testing method and is not investment advice.
Oct 30, 2026 · 4 min read
Oct 28, 2026 · 4 min read
Aug 18, 2026 · 8 min read
Jul 7, 2026 · 5 min read