I had a weekly RSI/OBV divergence indicator on my Bitcoin chart, and it looked incredible. Red marks sitting right on the tops, green marks right on the bottoms. The 2022 bottom near $15K, the 2025 top at $126K: both called.
Before using it to decide which grid bot to launch, I ran it through the same test I use for every signal. It didn't survive. The reason is worth knowing, because it applies to a lot of indicators you see on TradingView.
The mark is drawn in the past
A divergence needs a confirmed pivot: a low (or high) with N lower (or higher) candles on each side. This indicator uses 5. On a weekly chart that means the pivot only exists 5 weeks after it happened.
The script is honest about it: it only detects the divergence when it could have known. But then it draws the label on the pivot candle, five bars back. So when you scroll through history, every mark sits exactly on the extreme, and your eye reads it as "it called the top".
Here are the real Bitcoin signals, with the price at the pivot and the price on the day you could actually have acted:
| Bearish divergence | Price at the pivot | Price when confirmed |
|---|---|---|
| Jan 2025 top | $110,000 | $80,700 |
| Sep 2025 top | $125,900 | $94,200 |
| Oct 2025 top (RSI) | $126,200 | $86,800 |
The chart shows a perfect top call. Live, you got the message after a 25–30% drop.
In code, the whole difference is one line: the date you attach to the signal.
pivot_week = weekly.index[i]
# LB = 5: the week the bot could actually act
known_at = weekly.index[i + LB] + pd.Timedelta(days=7)
The test
I ported the indicator to Python (RSI(14) and OBV, 5/5 pivots, 5–60 bars between pivots, regular divergences only) and ran it on weekly candles for 8 pairs (BTC, ETH, SOL, XRP, DOGE, BNB, ADA, LINK) from 2020 to 2026.
Then I matched it against 4,846 simulated 30-day futures grid bots: for each start date, how much better a long grid did than a short one. The question: does a recently confirmed divergence predict that?
Same rule as every other signal I've tested, fixed before looking: it has to point the same way, with |ρ| ≥ 0.08, in three separate periods.
The results
First surprise: there are very few signals. 53 confirmed divergences in 6 years across 8 pairs. About one per pair per year. With that few, three or four memorable hits make it look like a system.
Correlation with (long − short), using the last divergence confirmed in the previous 4 weeks:
| 2021–23 | 2024–25 | 2025–26 | |
|---|---|---|---|
| Weekly divergence | −0.001 | +0.229 | +0.042 |
Great in one period, nothing in the other two. Fear & Greed did exactly the same in my earlier test.
Used directly (bullish → long grid, bearish → short grid, nothing → neutral), it beat always-neutral by +0.2 points per month, and did worse than neutral in 54 of 74 months. For comparison, the positioning signal I actually use gets +1.8 points per bot and wins 32 of 52 months.
One thing did look interesting: in the 4 weeks after a bearish divergence, long grids did 11.8 points worse than short. But those bots come from a handful of episodes, and in 2021–23 the effect isn't there. I wouldn't bet on it.
What I take from it
- Check where the label is drawn vs. when it's known. Any indicator built on pivots, zigzags or "confirmed" swings can look perfect on history for this reason alone. Scroll bar by bar with replay, or look at the timestamp of the alert, not the mark.
- Count the signals. If you can count them on two hands, the chart can't tell you whether it works.
- Split history. A signal that works in one market regime and vanishes in the next isn't a signal yet.
As a personal "maybe don't open a long grid this month" warning, a fresh bearish divergence is fine. As something to trade on, it didn't hold up.
I publish the signals that survive (and the ones that don't) here: https://gridreplay.com/en/validation/ — and the daily grid direction for BTC and ETH is at https://gridreplay.com/en/oracle/
Disclosure: I made both. The oracle page has a Pionex affiliate link, and a paid tier is planned for more pairs. Not financial advice.
If there's an indicator you think works for grid bots, leave it in the comments and I'll run it through the same test.
Top comments (0)