+702% on SOL. Where Is Jev in This Backtest?

Jev launched in September 2026, and the hype came quickly. Clones, curated lists, and trading bots appeared on GitHub. We decided to test the most tempting claim: can these bots earn money on historical markets after fees? A bot shows candles or an order book; the model chooses an action. Then come the numbers.

We collected 28 projects from awesome-jev and related lists. Their source code, pinned versions, and our test code are open on GitHub. One result is hard to ignore: over nearly three years, a trend rule returned +702.1% on SOL. The same rules lost 47.2% on LTC. The long test covered 15 crypto pairs. The question is what Jev itself contributed to those numbers, and what came from a conventional algorithm.

The headline number

The illustrations in this article are not charts of results. Verified returns appear in the table below.

On the longest uninterrupted SOL segment, the trend component of trading-bot-jev returned +702.1% after warm-up. Simply holding SOL over the same segment returned +312.2%. With higher fees and slippage, the algorithm's return fell to +632.0%; delaying entry by one minute produced +667.8%. The maximum drawdown in the base run was 34.1%.

The SOL return belongs to a particular historical period and a particular set of rules. On BTC, the base return of +146.2% only slightly exceeded the +139.6% from holding; once costs increased, it fell behind. On LTC, the same rules lost almost half the capital.

What we actually tested

For BTC, ETH, SOL, and 12 other pairs, we used available one-minute candles from 2023 to early 2026. Before calculating returns, we checked the minute-by-minute sequence: gaps split a series into separate tests, and capital resets at the start of each new segment. The first 400 four-hour candles warm up the indicators and do not count toward returns.

We fixed the list of 15 pairs before looking at returns. That left 23 separate pair segments and one portfolio of seven pairs without gaps. Each went through three scenarios: standard costs, higher costs, and entry one minute after the next candle opened. That is 72 runs in total. A signal uses an already closed four-hour candle; the base execution occurs at the next candle's open.

The long test uses Binance USDT perpetual futures one-minute candles. Binance documentation describes this candle type. The original bot was designed for Bitvavo spot trading in USDC pairs. We used futures candles as a price proxy. Simulated strategy trades carried a 0.05% fee per side and four basis points of slippage. Funding payments, the exact spread, and actual Bitvavo fills were excluded. The table therefore does not show what one could have earned on the original exchange.

The winners change with the calendar

For each pair, the table shows its longest uninterrupted segment. We selected that segment by duration, before looking at profit. The period refers to the source candles; returns start only after the 400-candle warm-up. “Hold” marks a purchased position at the closing price, with one entry fee but no slippage or exit fee. It is a comparison point, not a complete estimate of the proceeds from selling. “Stress” raises the strategy fee to 0.1% per side and slippage to 10 basis points; “+1 min” moves the entry price forward one minute at the original costs.

Seven pairs—ETH, SOL, ADA, AVAX, LINK, DOT, and UNI—remained profitable and beat holding even with higher costs and a one-minute delay. This list was obtained after the test. Selecting only these pairs now would use knowledge of the future again, this time at the researcher level.

Among the seven pairs with an almost continuous three-year history, the algorithm beat holding on just one in 2024. It did so on four in 2025, but only two had a positive absolute return. SOL lagged holding in 2023 and 2024; most of its relative advantage arose when the market fell in 2025.

Over the entire active period, the seven-pair portfolio returned +189.0% versus +129.6% from holding, with a maximum drawdown of 37.2% versus 55.4%. Yet 2024 was weaker than holding here too: +45% versus +71%. In 2025, the algorithm lost about 6% while holding lost about 17%. At times it weathered a decline better; that is different from earning consistently.

Could the bot peek at tomorrow's candle?

We checked the original backtest engine and the input series. The decision for candle i receives closes through i, inclusive. The trade occurs at the open of i+1. We then truncated the history at different points 20 times or changed all subsequent prices sharply. Trades and equity before the boundary remained byte-for-byte identical. We found no direct reading of future candles in the trading path we checked.

We did find a less visible bug: the original engine labeled an equity point with the next candle's opening time, although it valued equity at that candle's close. We corrected the label in the research wrapper; trades and returns did not change. We covered how such shifts corrupt tests in our look-ahead bias analysis.

Still, the original project's configuration says explicitly that the EMA 200 and the 1% band were selected using a three-year backtest. The author had already seen the history when choosing the rules. Our test cannot turn it into an independent holdout sample. A one-minute delay checks sensitivity to entry timing, but does not replace quotes and fills from the actual exchange.

Where is Jev in those returns?

The news input to Jev was empty in the long test: we lack an archive of news with publication timestamps and saved model responses to that news. The +702.1% on SOL and +189.0% for the portfolio belong to the trend rule; the model's contribution to them was not tested.

We also tested Jev itself where the data allowed. Across three separate two-hour windows of Bybit order book data, it made 21 decisions and had 10 simulated fills. The sum of results after trading fees, with capital reset for each window, was +3.91 USD. A simple momentum rule on the same windows made +14.33 USD. At this bot's standard 400 ms response limit, the first window had no fills at all. Three two-hour windows are too little to judge profitability: order book states were recorded once a minute, and the order queue and trades within each minute were not modeled.

Another project asked Jev to predict the next day's direction. On 1,027 eligible historical dates, its accuracy was 35.93%. A baseline that used only already known outcomes each day achieved 42.55%. Here we tested predictions; we did not simulate orders. For dates from 2023 onward, we called the current model in 2026. We do not have historical snapshots of it or a verified training cutoff. Jev was introduced publicly in September 2026; the model might have learned about later events during training. This run does not prove it would have made the same predictions in the past.

When would we call this a profitable strategy?

On particular pairs and periods, fixed trend rules made money in simulation. We have no confirmation that ready-made Jev bots earn reliably in live trading. We cannot credit the model with a result from a test in which it made no decisions.

The next experiment must begin before new results appear: lock the parameters and the pair-selection rule in advance, then collect trades on a fresh segment of Bitvavo's USDC spot market with actual quotes and costs. Testing the news filter requires news with publication times and Jev's responses saved at the time. The same trade candidates must be compared with and without the filter. The next report should show trades made after those rules were fixed.

We are testing open alternatives to Jev separately on the same trading task. Benchmark code, reproduction instructions, and results as runs finish are available on GitHub; the table by model and crypto pair shows what has been measured so far.