How we tested
TickTurn scored its own predictions. The short version: we cannot claim it has skill. Below is how the scoring was done, what came out, and what the result cannot say.
1. What was scored
We showed the model only the bars up to a given day, asked “does the next trading day close higher?”, then checked the answer against what actually happened. Then we moved one day forward and did it again.
| Period | 3 July 2021 – 5 August 2026 |
|---|---|
| Symbols | 25 — 10 Korean stocks (Samsung Electronics, SK hynix, NAVER, Kakao, Hyundai Motor, LG Chem, Celltrion, POSCO Holdings, KB Financial, Samsung Biologics), 10 US stocks (AAPL, MSFT, NVDA, TSLA, AMZN, GOOGL, META, JPM, KO, JNJ), 5 coins (BTC, ETH, XRP, SOL, DOGE) |
| Readings in total | 60,842 |
| Held out for scoring | The most recent 30% by time: 14 March 2025 – 5 August 2026. Of those we count the 17,971 where the model actually picked a side (a score of exactly 50 is treated as no call and dropped). |
| What counts as “up” | Next close > today’s close. An unchanged close counts as not up. |
| “Next trading day” | The Korean exchange calendar for Korean stocks, the US local calendar for US stocks, and the next day for crypto (Upbit’s daily bar closes at 09:00 KST). Ad-hoc market holidays are not modelled; only weekends are skipped. |
| Bars fed to the indicators | Closed daily bars only. A bar still in progress is excluded. |
2. What it was measured against — this is the part that matters
At first we used the share of up days in the period as the baseline. That is the wrong yardstick. In a stretch where only 49% of days rise, saying “down” every single time is right 51% of the time — so a model that happened to say “down” a lot in a falling market looks like it beat the baseline without having any skill at all.
The right baseline is “guessing at random, in the same proportion the model called it”. Measured that way, a proud +1.26%p shrank to +0.51%p.
| Model accuracy | 50.50% |
|---|---|
| Old baseline (share of up days) | 49.24% → +1.26%p (a number we should not use) |
| Random, same call proportion | 49.99% → +0.51%p (the honest number) |
| Margin of error | ±0.37%p |
+0.51%p is 1.4× the margin of error. The usual bar for claiming an effect is at least 2×. So: we cannot claim it has skill.
3. Split by asset class
| Korean stocks | +0.09%p — effectively zero |
|---|---|
| US stocks | −0.58%p — worse than the baseline |
| Crypto | +1.32%p |
Outside crypto there is nothing there. And splitting the sample shrinks it, which widens the margin of error — these three numbers may themselves be noise.
4. The size of the score means nothing
We sorted the scored period by score, cut it into ten equal buckets, and checked whether the confident buckets actually rose more often.
| Bucket 1 · score 0–17% | rose 48.4% of the time |
|---|---|
| Bucket 5 · score 38–50% | 48.2% |
| Bucket 9 · score 76–83% | 51.7% |
| Bucket 10 · score 83–100% | 48.2% |
The bucket that shouted “up” loudest and the one that shouted “down” loudest rose at the same rate. A higher score does not mean a bigger move. The number on the screen is not a probability — it is only how far the 21 indicators lean one way.
5. It is not an artefact of directional bias
The 9,068 calls of “up” beat the period’s share of up days by +0.51%p, and the 8,903 calls of “down” beat the share of down days by +0.52%p. The two sides are symmetric, so this is not an illusion created by leaning one way. It is simply very small.
6. What this test cannot tell you
- The 21 indicators are not independent. Moving averages, MACD and the disparity index all come off the same average; RSI, Stochastics and CCI measure the same overheating in different ways. Similar signals can be counted several times and inflate the score.
- 25 symbols is few, and all of them are large and heavily traded. Nothing here says the same holds for thin or wildly volatile names.
- No fees, taxes or slippage. We counted whether the direction was right, not what would have been left after trading it.
- One window only. This is the result for March 2025 – August 2026; another window is not guaranteed to look like it.
- Ad-hoc market holidays are not modelled. Only weekends are skipped, so “the next trading day” can be off by one around a holiday.
7. So what is this tool
Not a machine that predicts the future, but a summary tool that reads 21 technical indicators at once. Its use is turning a job you would do indicator by indicator into a single screen, and the numbers above do not support any claim beyond that.
Nothing here is a recommendation to trade, and no return is guaranteed. If you have questions about the calculation or the test, write to us on the contact page.
← Back to home