Every signal is a number.
Every number shows its work.
Analytica Trading layers a cross-sectional factor model, alternative data, a relationship graph and a macro-regime classifier on top of the technical signals every tool has. Nothing is a black box: each signal carries its rule, its weight, its source, and its exact contribution — metered on one −100 to +100 scale, never reduced to a bare “buy”.
How a signal becomes a decision
Two independent engines score every name. Their blend is fixed, published, and decomposes exactly — the combined score is always the sum of its parts.
Technical rules
Five bounded point rules on daily candles, computed across 1M / 3M / 1Y so timeframe agreement is visible.
Cross-sectional factors
Twelve factors z-scored against the whole universe, tilted by the macro regime. Relative ranking, not absolute thresholds.
One −100…+100 meter. Worded as a lean (“leaning bearish”), never a command. The same blend ranks the CIO book.
Direction-aware trade plan
Long, short, or stand aside — with entry, stop, target, reward:risk, and how much of the move is already done.
The full decomposition is public inside the product: every stock has a one-page evidence report where each signal appears with its raw metric, rule, weight, source and exact contribution — and the contributions sum to the combined score to the decimal.
The analytics, itemized
Every live signal with its real parameter. These are the numbers running in production today, not a brochure list.
Technical point rules (Engine A)
| RSI (14) | ±25 | <30 oversold bullish · >70 overbought bearish · lean bands +8/−6 |
| MACD histogram | ±15 | above the signal line bullish, below bearish |
| SMA 20/50 stack | ±20 | rising stack uptrend · falling stack downtrend · single-line ±6 |
| Volume conviction | ±6 | >1.5× the 20-day average amplifies the prevailing direction |
| Headline sentiment | ±12 | lexicon-scored recent news, positive vs negative skew |
Stock factor model (Engine B — weights)
Crypto factor model
Top ~100 coins, stablecoins and wrapped assets filtered out. Plus Binance perp positioning: funding, open interest, long/short crowd.
Regime classifier
S&P 50/200 trend + 21-day realized volatility + universe breadth → risk-on / risk-off / neutral. The regime does not pick stocks; it re-weights the factor model (momentum up to ×1.5 in clean trends, low-vol up to ×1.6 in stress) and gates the bots’ long entries.
On top: a curated relationship graph(suppliers, customers, cost inputs, peers) that flags moves a name hasn’t priced yet, and event overlays — insider clusters, earnings drift, natural-disaster exposure.
The Trading Arena
Strategy claims are cheap. The arena makes them measurable: autonomous paper bots run the same risk framework over different signal blends, so which signal pays is an experiment, not an opinion.
35 bots, 2 asset classes
17 stock + 18 crypto virtual traders, each a controlled experiment: one signal profile, fixed risk rules, long and short, no leverage.
Realistic fills
Every simulated fill pays slippage (5 bps stocks, 10 bps crypto) in the adverse direction. Fractional sizing in crypto. Busted books liquidate and halt.
Benchmark controls
A SPY buy-and-hold bot and a BTC buy-and-hold bot run inside the fleets — the null hypothesis every strategy must visibly beat.
The Champion loop
A meta-bot adopts the best live strategy only after enough evidence (8+ daily points, Sharpe margin, cooldown) and only if a point-in-time backtest confirms the signal's historical edge.
Validation we publish, including the failures
Most retail tools show you an optimized backtest. We show the out-of-sample number next to it — that difference is the point.
Price-only signals look brilliant in-sample and die out-of-sample. We publish that collapse instead of shipping the optimized weights — it is the quantitative argument for the alt-data stack above, and the reason the optimizer now runs walk-forward (expanding training windows, stitched unseen test blocks, weights shrunk toward equal) rather than a single lucky split.
Point-in-time snapshots
Every day the system records what it knew that day — factors and prices, stocks and crypto. Alt-data becomes backtestable without look-ahead bias, honestly.
No silent failures
A public health probe watches the fleets; monitoring alerts the team the half-hour anything stops ticking.
Decision support, not advice
Outputs are metered leans with visible evidence, built for a professional to interrogate — not trade commands.
Data sources
Deliberately built on free and public tiers first — the architecture proves out before a dollar goes to premium feeds. Every feed is cached and paced to its provider's limits.
| Provider | Used for | Tier |
|---|---|---|
| Twelve Data | equity candles and charts | free tier, paced |
| Finnhub | quotes, news, fundamentals, ratings, earnings, insider filings | free tier, paced |
| SEC EDGAR | 13F smart-money holdings | public |
| StockTwits | retail message tone | public |
| ApeWisdom | Reddit mention velocity | public |
| CoinGecko | crypto markets and candles | keyless |
| Binance Futures | perp funding, open interest, long/short crowd | keyless |
| USGS / NASA EONET | earthquakes and storms vs exposure map | public |
| Claude / DeepSeek | AI analyst synthesis, cached 6h | budget-capped |
The signal roadmap
65 further signals are catalogued in-product with a cost badge each — a costed growth path from free public data to the institutional feed set, prioritized by measured impact once point-in-time history can judge them.