Recommended Free Tools
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
In one test, Jev did not reliably predict whether the next Bitcoin five-minute candle would close up. Jev’s area under the receiver operating characteristic curve (AUC) was 0.486 using numeric indicators and 0.494 using text descriptions—both near, and slightly below, the 0.500 no-skill reference. The more striking finding was directional: Jev tended to read stretched price and momentum indicators as signs of continued movement, while the tested BTCUSDT data more often showed a small reversal.
Those results concern Jev version 1.13.0, Binance BTCUSDT five-minute candles, and price-derived inputs. They are a narrow experiment, not proof about every Jev version, market, or task.
What did the Jev Bitcoin test measure?
The test asked a specific question: after a BTCUSDT five-minute candle closed, would the next candle close above its own open? That is different from forecasting Bitcoin’s longer-term price, predicting a large move, or deciding whether to buy or sell after fees.
According to the report by truongxx, published September 27, 2026, the underlying data came from Binance’s public API. The reported training period ran from January 1, 2022, through December 31, 2024, and included 314,552 candles. A validation period covering July through December 2024 was used to make choices. The test period ran from January 1, 2025, through September 24, 2026, and included 181,290 candles. The author calculated 51 indicators, including moving averages, RSI, MACD, ATR, Bollinger measures, candle shape, streaks, and 15-minute and one-hour candle data.
For Jev’s comparison, the author selected 12 indicators using the validation window, then tested Jev version jev-1.13.0 on 1,500 randomly selected unseen candles. Jev received the indicators either as numbers or as plain-text descriptions. The author says the requests were blinded to symbol, dates, and prices. Each also included control questions whose answers were contained in the supplied data, such as whether RSI was above 50. Jev answered those controls correctly 99.8% of the time, which supports the view that the main result was not simply an inability to read the inputs. See the author’s report for the experiment details.
How did Jev compare with the baselines?
Accuracy measures how often a predicted up-or-down label was correct. AUC measures how well a model ranks upward outcomes above downward ones across possible classification thresholds; 0.500 is the no-skill reference. The two measures are not interchangeable.
| Model or rule | Accuracy | AUC | Evaluation |
|---|---|---|---|
| Always predict the base rate | 49.9% | 0.500 | Full 181,290-candle test set; reported by the article author |
| Continue the last candle’s direction | 49.5% | 0.495 | Full test set; reported by the article author |
| Reverse the last candle’s direction | 50.6% | 0.506 | Full test set; reported by the article author |
| Logistic regression, 51 indicators | 51.5% | 0.522 (95% CI 0.520–0.525) | Full test set; reported by the article author |
| Gradient boosting, 51 indicators | 51.9% | 0.527 (95% CI 0.524–0.530) | Full test set; reported by the article author |
| Gradient boosting, selected 12 indicators | 51.7% | 0.526 (95% CI 0.524–0.530) | Full test set; reported by the article author |
The more direct comparison used the same 1,500 unseen candles for Jev and the selected-indicator boosting model:
Rank #2
| Model | Accuracy | AUC (95% CI) | Control questions |
|---|---|---|---|
| Gradient boosting, selected indicators | 51.6% | 0.522 (0.493–0.550) | Not applicable |
| Jev, numeric indicators | 48.1% | 0.486 (0.459–0.516) | 99.8% correct |
| Jev, text descriptions | 49.3% | 0.494 (0.463–0.524) | 99.8% and 99.7% correct, respectively, as reported for the two formats |
All three AUC confidence intervals in this small comparison overlap 0.500. The figures therefore do not establish that boosting was definitively better, or that Jev was definitively worse, on this sample. They do show that Jev did not demonstrate above-chance AUC in the reported test.
What does “read the indicators backwards” mean?
The report’s “backwards” description is about the direction of the associations, not a claim that every individual prediction was the exact opposite of the outcome. In the tested data, greater distance above EMA9 was associated with a lower chance of the next candle closing up. Jev’s predictions moved in the other direction.
- For distance from EMA9, the reported rank correlation was −0.80 for boosting, +0.83 for numeric Jev, and +0.72 for text Jev.
- For the move over the last three candles, it was −0.66 for boosting, +0.76 for numeric Jev, and +0.70 for text Jev.
- For same-color candle streaks, it was −0.53 for boosting, +0.65 for numeric Jev, and +0.87 for text Jev.
- For RSI(14), it was −0.76 for boosting, +0.77 for numeric Jev, and +0.51 for text Jev.
The author also reports that the two models’ predictions correlated at −0.71. As an example from the test data, 52.8% of next candles closed up when price was far below EMA9, compared with 46.9% when it was far above. Jev’s average predicted probability reportedly rose from 41.6% in the far-below group to 54.4% in the far-above group.
The proposed explanation is that Jev applied a familiar momentum-style reading—elevated price, RSI, or a run of green candles suggests further upward movement—while this particular five-minute BTC sample showed a slight tendency to pull back after stretched moves. That is an interpretation of the tested dataset and setup, not a general rule that Bitcoin reverses whenever an indicator looks overbought.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Would these results support a trading strategy?
No. The author reports that the selected-indicator boosting model’s most confident 10% of calls returned 0.74 basis points per trade before costs. The report compares that with an estimated 15–20 basis points for a Binance spot round trip. These are figures for the study’s setup, not a current fee quote or a forecast for other markets; in that comparison, the reported pre-cost signal was well below the stated trading-cost estimate.
The author also tried combining Jev with boosting in three ways: replacing boosting with Jev, trading only when Jev agreed with boosting, and blending their probabilities. Replacing the model performed worse. Agreement filtering underperformed allowing boosting to select its own most confident calls, and blending did not materially improve results. The fitted Jev weights were −0.09 for numeric inputs and +0.16 for text inputs. In one comparison, numeric Jev agreed with only 6 of the 300 most confident boosting calls and objected to 294.
Those tests are a reason to be cautious about using Jev as an order gate in this setup; they do not show that every more complete trading system or possible entry filter would fail.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the experiment does not establish
The result is limited to Jev 1.13.0, Binance BTCUSDT, five-minute candles, a 1,500-candle Jev sample per format, and indicators derived from price. The author did not test news, filings, order-book context, or prediction-market data. A newer Jev version, another asset or timeframe, or a task with information beyond price could produce different results.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →The author notes that prompting Jev with an explicit reversal rule might change its answers. But following a supplied rule would not, by itself, demonstrate that Jev independently discovered the pattern. The report is a late-September 2026 snapshot and points to public code as a way to rerun the experiment; the reported test has not been independently reproduced here.
How to evaluate an AI claim about trading
Before treating a model’s market prediction as useful, check whether the evidence answers these questions:
- What is the baseline? Compare the model with a simple rule and a conventional model on the same task, not just with a claim of impressive accuracy.
- Was the test genuinely out of sample? Data used to select indicators, tune prompts, or choose thresholds should not also serve as the final performance test.
- Are uncertainty and costs reported? Look for confidence intervals as well as fees, spread, and slippage; a small metric advantage may not survive execution costs.
- Did the model read its inputs correctly? Control questions can help distinguish a prediction problem from a parsing or formatting failure.
- What happens when the model is uncertain or wrong? A usable system needs explicit failure behavior, not just a confident probability.
The author’s broader assessment is that structured-data tasks such as classification, labeling, or checking a trading journal against stated rules may be a more natural fit for Jev than forecasting a future price move from indicators alone. The examples do not prove that Jev—or AI systems generally—will perform well on all such tasks. For a fair comparison, keep the task and unseen data consistent, report uncertainty and input-reading controls, and examine results after costs.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitches
