Claude+ fine-tuned with Tuatara Vectors: Performance Returns

A leading edge gives you alpha. Headline→asset signal on US equities at a 2-session hold (~2 trading days, long basket). Each model runs the same news flow; gross is the unhedged directional return, vs SPY is the excess return over the S&P 500 over the same window.

Get your own Tuatara model →

Claude+ with Tuatara performance returns comparison on US equities, 2-session hold, gross and vs-SPY returns for Tuatara, Claude, Combined and MiniLM

On equities, the Tuatara vector signal leads outright — on return, win rate, and risk-adjusted Sharpe. Forced long over a 2-session hold, Tuatara returns +2.28% gross / +1.94% over SPY (win 68%), well ahead of Claude +1.21% / +0.86% (win 66%), and it does so at lower volatility too, so its per-trade Sharpe is 0.70 vs Claude's 0.27. The Combined (Tuatara ∪ Claude) book is best on a risk-adjusted basis: +1.80% over SPY at Sharpe 0.79, win 72%. The free on-box MiniLM embedder lands far behind (+0.85% gross / +0.51% over SPY, win 54%, Sharpe 0.20) — on the equity universe the discriminative Tuatara vectors clearly beat a generic sentence embedder. All legs are measured in a single run on the identical selected headlines (top-K 12), so the only thing that differs is each model's asset picks.

2-session hold — by model (US equities)

Model	Gross %	vs SPY %	Win %	Sharpe / trade	n
Tuatara (vector)	+2.28	+1.94	67.8	0.70	149
Claude (LLM reasoning)	+1.21	+0.86	66.3	0.27	92
Combined (Tuatara ∪ Claude)	+2.15	+1.80	72.5	0.79	149
MiniLM (on-box, free baseline)	+0.85	+0.51	54.0	0.20	150

Sharpe is per-trade (mean ÷ stdev). Claude trades only the headlines it picks a basket for (n=92); Tuatara/Combined/MiniLM cover the full selected set.

Parameters used

Asset class / namespace	US equities · `CMDB-nasdaq-v07` (asset_deep_discovery, finnhub-us)
Headlines (sample)	150 selected → 149 priceable, entry at next-session open after the headline
Direction	forced long (`--force-side long`) — every non-empty basket traded long
Hold	2 trading sessions (~2 days, `--hold 2`); exit = close +2 sessions
Basket size (top-K)	12 (`--top-k 12`)
Score filter	Tuatara min score 0.04 (`--min-tuatara-score 0.04`), top-K by score
Universe	all priceable tickers (27,848 with a bar; embed universe = 15,771 with a ≥40-char description)
Round-trip cost	0.15% (`--cost 0.0015`)
Split/artifact guard	per-leg move capped at ±3.0 (`--max-leg-move 3.0`)
Price source	`historical_stock_prices` (daily OHLC, weekday sessions)
Benchmark	SPY (S&P 500 ETF)
Gross	unhedged equal-weight long-basket return
vs SPY	basket return − SPY return over the identical window (excess return)
Tuatara	vector engine, query = full article body (keyword-condensed), top-12 by score
Claude	entitlement engine, model `claude-opus-4-8` (baskets cached, query = title)
MiniLM	on-box ONNX `all-MiniLM-L6-v2` (384-dim); cosine over the 15,771-stock description universe

All legs are measured in a single run on the identical selected headlines; only the asset picks differ. Command: backtest_tuatara_vs_claude_stocks.py --sample 150 --hold 2 --force-side long --cache claude_stock_baskets_cache --embed-baskets embed_stock_baskets_minilm --embed-label MiniLM

Method

Asset class US equities: prices from historical_stock_prices (daily OHLC), trading only on weekday sessions ~09:30–16:00 ET. Baskets from each model over the same headline corpus; Tuatara from the vector engine asset_deep_discovery (CMDB-nasdaq-v07), top-12 by score. Long basket, exit at the close +2 sessions: per leg close(t₀+2)/open(t₀+1) − 1, equal-weight, split/artifact-guarded, minus round-trip cost. vs SPY subtracts the S&P 500's return over the identical window. Sharpe is per-trade (mean ÷ stdev).

The embedding harness is model-agnostic: the generator build_stock_embed_baskets.py embeds each stock's profile (name + ontology + description from the trading_vehicle_instances corpus that also trains Tuatara's nasdaq-v07 model) and each headline with the chosen model, cosine-ranks the top-K, and emits the identical proofs schema the backtest reads — so any additional embedding model (OpenAI text-embedding-3, Google Gemini, Cohere) drops into this same comparison with one command. The free on-box all-MiniLM-L6-v2 embedder shown here is the baseline.

Download chart (PNG) Crypto hourly return curve →

Cymetica · Tuatara vector research · signal/research only, not investment advice · cymetica.com