patternfetch vs raw OHLCV & candle APIs

They are not really competitors so much as different shapes of the same underlying data. A raw API is a firehose you process yourself. patternfetch is a pre-chewed summary an LLM can read directly. The right choice depends on whether you or a model is the consumer.

Side-by-side

"Higher/lower hallucination risk" is relative, not a benchmark: the more raw arithmetic an LLM has to do on long number arrays, the more it tends to slip — which is the failure mode the interpreted brief is designed to avoid (see the methodology). We've tried to be honest in both directions: patternfetch wins on agent-readiness and token cost; raw APIs win clearly on granularity and control. "Generic crypto data API" varies a lot by vendor, so those cells say "varies" where a blanket claim would be unfair.

When to use raw OHLCV instead

Be candid: there are jobs where a raw candle/OHLCV API is simply the right tool, and patternfetch is not.

• Backtesting your own strategy. You want full, deterministic history at native resolution to replay your logic bar-by-bar. A raw export beats a digested brief here.
• Custom indicators. If you've built proprietary signals and want to compute them yourself from first principles, you need the raw inputs, not someone else's interpretation.
• Sub-minute granularity. Tick-level or sub-minute data for microstructure work is outside what a market-state brief is meant to carry.
• Asset classes we don't cover. patternfetch covers US stocks, ETFs and crypto spot in one schema. If you need FX, futures or options, a broader raw data vendor is the fit.

In all of these, you are the consumer and you want maximum control. That's the firehose case.

When patternfetch wins

The flip side: when an LLM or agent is the consumer, the digested brief is usually the better shape.

• LLM / agent context. The brief is already framed as patterns, levels, regime and a summary — exactly the abstractions a model reasons over, with no preprocessing step.
• Token budgets. A few hundred tokens instead of a multi-thousand-token candle array per ticker. Across many symbols or many turns, that's the difference between affordable and not.
• Avoiding numeric hallucination. Interpreted indicator state and a one-line summary mean the model rarely has to do arithmetic on raw numbers — where LLMs are least reliable.
• Fast integration via MCP. Add the MCP server, click Authorize, done — no SDK, no indicator library, no data plumbing.
• No infra. No streaming pipeline, no indicator computation, no candle storage. One call returns the whole market state.

The same data, two shapes

[
  [1718000000000, 60125.4, 60480.0,
   59890.1, 60310.7, 1284.5],
  [1718014400000, 60310.7, 60720.2,
   60180.0, 60655.9, 1102.8],
  … 198 more rows …
]
// then: you compute RSI, EMA, ATR,
// find patterns, cluster S/R, label
// the regime — and feed it all to the
// model as thousands of tokens.

{
  "codec": { "rows":"60125.4,60480,…",
             "sax":"dcefdcbe","precision":1 },
  "analysis": {
    "patterns":[{"name":"double_bottom",
                 "confidence":0.86}],
    "levels":{ "support":[{"price":59820.4}],
               "resistance":[{"price":63450.8}] },
    "regime":{ "trend":"up","volPct":2.13 },
    "indicators":{ "rsi":{"v":58.3,"state":"neutral"} },
    "nl":"BTC/USDT: uptrend (moderate),
          RSI 58.3 (neutral), double_bottom (conf 0.86, n=18, hist 51% over 10b)."
  }
}

Note the brief still carries codec.rows — the underlying OHLCV numbers — so you're never locked out of the raw data. You just don't have to spend tokens (or a model's arithmetic) on it unless you want to. See it live →

Pricing snapshot

The honesty layer: every pattern ships its base rate

Most pattern and signal APIs hand a model a confident-looking "double_bottom", 0.86 and let it quote that blind. patternfetch attaches an evidence block to every detected pattern — the backtested historical hit-rate for that exact pattern, timeframe and confidence band, the pattern-free baseline for the same market and horizon, and a cluster-robust interval on the difference — measured with no lookahead (chart patterns are scored only from the bar they become knowable). A hit rate on its own cannot be read: a drifting market moves it away from 50% with no pattern involved. Measured against the baseline instead, three of the 105 buckets we tested clear an uncorrected interval — fewer than the ~5.3 that chance alone would produce, and all three are crypto. The only one that lands above its baseline (double_top 1d, +9.78pp) rests on n=149 and does not survive correcting for having tested 105 buckets; on US stocks & ETFs it is 0 of 60. The evidence block reports exactly that, instead of laundering a shape score into a confident call. Impersonal market data, not advice. See the method →

FAQ