The idea in one paragraph
Nobody can see tomorrow's gold price, but history can tell us how far gold usually moves in conditions like today's. Each morning the program describes the market with five numbers, finds the days in our 16-year history that looked most alike, and counts what happened over the next one and two trading days. That count becomes the odds, the likely range and the chance of touching a level.
The five numbers that describe a day
| Measure | What it captures |
|---|---|
| Volatility rank | How wild the last 20 days were compared with the previous two years |
| 20-day trend | How far price sits above or below its 20-day average, in units of a normal daily move |
| 5-day trend | The move over the last five days, in the same units |
| Last-day move | How big yesterday's move was compared with normal |
| Range position | Where the last close sits inside the last 10 days' high-low range |
From similar days to odds
- Find the 250 most similar past days, using only days whose outcome was already known at the time.
- Look at what followed: the close one and two trading days later, and the highest and lowest prices in between.
- Count. The share that ended up, flat or down gives the direction odds; the spread of closes gives the range; the highs and lows give the chance of touching a level.
- Pull toward the long-run average. With a limited number of similar days, raw counts can look more certain than they are, so the odds are blended with the all-history average.
- Log it. The forecast is saved before the day starts and scored when the day closes.
What the back-test found
Before trusting the method, we ran it over 3,595 past trading days, each time using only what was known that morning. Three results matter:
| Test | Next day | Next two days |
|---|---|---|
| Direction skill vs the long-run average | 0.16% | 0.27% |
| 50% range actually held | 49% | 49% |
| 80% range actually held | 78% | 78% |
What the model does not know
- Scheduled news. It does not know that a jobs report or an interest-rate decision is due. See how to read news data.
- Positioning, sentiment and geopolitics. None of these are inputs.
- A regime it has never seen. In a market unlike anything in the data, the similar days are not very similar. The outlook flags this as an unusual setup.
- Costs. Spreads and slippage are not modelled, because this is not a trading system.
How we keep ourselves honest
- Each forecast is saved before the day starts and cannot be overwritten by our software.
- Every forecast is scored automatically and shown on the public scorecard, wrong ones included.
- No number in any post is typed by hand: it comes from code. A post is held back if its checks fail.
- The AI only writes a short note around the numbers, and the note is thrown away if it contains a number it was not given.
Frequently asked questions
Is this a trading signal?
No. It is educational research. The direction odds are close to the long-run average on most days, and we say so every time.
How many past days does it use?
It looks at about 4,000 trading days since 2010 and uses the 250 most similar to today for each forecast.
Why not use machine learning?
We tested a simple, transparent method first. It found almost no direction edge, which is itself the lesson: more complexity would mostly add overfitting.
Sources and method
- Daily XAUUSD bars from our own MetaTrader 5 history, 2010 to today. Back-test: 3,595 days, k=250, scored with the Brier score against the long-run average.