Renaissance Technologies doesn't use fundamental analysis. They feed raw market data into models and look for statistical patterns that humans can't see. That idea has been rattling around my head since I started testing TypeSafe's Jev — a "System One" model that returns probabilities in 175ms for $0.00007 per call. Jev doesn't see images, but candlestick data is just numbers: open, high, low, close, volume. What if you fed it price charts as text and asked it to read the tape?
Backtest v1: just ask
I downloaded a month of 5-minute BTC/USDT candles from Binance — 8,640 bars. For each bar, I sent Jev the previous 20 candles plus RSI, MACD, and volume stats, and asked a single yes/no question: will price be higher in 5 minutes?
Jev uses a "noul" primitive for yes/no — it returns a probability between 0 and 1. A well-calibrated model that genuinely doesn't know should return 0.5.
That's exactly what happened. Jev answered 0.5 for 82% of the 500 predictions. It was honest: it said "I don't know" over and over. The remaining 18% were clustered tightly around 0.5, with a slight lean toward "down." Overall accuracy: 50.6%. A coin flip.
Backtest v2: force a pick
Maybe Jev needs more information, or a stronger push. For v2, I expanded the state to 30 candles with percentage changes, body colors, wick lengths, and trend analysis. I also switched from yes/no to forced choice: up, down, or flat. And I added a second question about momentum: strong bearish to strong bullish on a five-point scale.
Forcing a choice didn't add signal — it added bias. Jev predicted "down" 63% of the time even though BTC went up 54% of the time in this window. There's a bearish lean baked in somewhere, probably an artifact of the training data distribution.
The inverted signal
The momentum question produced the strangest result. I asked Jev to rate the last five candles from "strong bearish" to "strong bullish." Then I checked what actually happened next.
When Jev said "mild bearish," BTC went up 64% of the time. When it said "mild bullish," BTC went up only 46% of the time. The signal is inverted. If you did the opposite of what Jev recommended, you'd have a 10-point edge over the base rate on the "mild bearish" bucket.
But I wouldn't trade on that. The sample is 500 predictions over one month. The "mild bearish" bucket has 205 entries and the "strong bullish" has only 38. This is the kind of pattern that looks great in a backtest and evaporates in production.
Why it doesn't work
Jev is a text model trained for calibrated decisions on natural language. OHLCV data is not natural language. Jev has no concept of mean reversion, support levels, or volume profiles. It's reading numbers and pattern-matching against training data that wasn't price charts.
The honest v1 result — 0.5, "I don't know," 82% of the time — is actually the correct answer. A language model that hasn't been trained on market microstructure should not have an opinion about whether BTC will be higher in 5 minutes. The fact that v1 said so is a point in Jev's favor, not against it.
The v2 result is what happens when you force an opinion from a model that doesn't have one. You don't get signal. You get the model's priors, which in this case include a bearish lean and an inverted momentum read. Asking it to trade is like asking a spell-checker to diagnose an engine knock from audio frequencies written out as text.
All code and data are at github.com/noreff/jev-signal.