Recommended usage: treat the graph as research, the backtest as proof
Corrada is a research copilot, not a strategy factory. Find tickers and relations that actually exist on the graph, write one testable hypothesis, then walk-forward — do not treat a snapshot z as a trade signal.
Corrada is a macro and theme relationship graph: who moves with whom, who leads, which relative-strength legs look stretched. It turns “I want oil, semis, rotation” into candidates you can check — not a strategy you can ship.
Treat the graph as the alpha-search layer, and walk-forward backtests as the proof layer. The one-sentence hypothesis in between is the thing you actually own.
This is a research workflow, not investment advice.
Four-step usage
1. Start from a theme. Do not invent tickers.
After you connect Corrada in Cursor or Claude, name the theme you care about: oil, lumber, semiconductors, risk appetite. Let the agent search the graph for matching ETFs, stocks, relative strength (RS), spreads, and narrow-industry PCA residuals.
Do not hard-code example symbols, and do not drop “looks similar” names into a basket. What the graph should return:
- An equation (for example
RS:XLE_SPY) - A ticker
- A relation type (correlation, cointegration, lead–lag, …)
- Why it is worth a look
Leave numeric weights for the validation step. Search results are brainstorming and backtest hypotheses, not trade instructions.
2. Write one sentence of hypothesis
Before you download prices, lock in a single sentence:
- What the signal is: relative strength, a spread residual, or a PCA residual
- What you trade: an ETF, a spread, or the residual itself
- Entry and exit: quantile, z threshold, how long you hold
- When you stand aside: regime (rates, oil/gold, risk appetite)
Without that sentence, the backtest is just fitting noise.
3. Download prices, recompute signal history yourself, then walk-forward
The z on the graph is a snapshot, not a history. Recompute the series you actually intend to trade from prices, then:
- Search parameters in-sample
- Score them out-of-sample
- Pick the best settings
- Run the final backtest with those settings
Do not report a pretty full-sample curve with fixed parameters. Overlay buy-and-hold of the same instrument on the equity curve so you can compare.
Data can come from Corrada daily bars, yFinance, or your own files. Pick one source and stay with it.
4. Use the six toolkit layers as a checklist, not as required indicators
On every strategy change, walk through: regime, support/resistance, volatility, volume, VWAP, TWAP. The graph mainly helps you find regime proxies (for example oil vs equities, gold vs real rates). Volume and VWAP are not on the graph today — mark them N/A when there is no data; do not pretend you passed the check.
When results look weak (drawdown too deep, Sharpe low, or worse than buy-and-hold), read the trade log and the failure windows first. Then decide whether to change the signal, change the filter, or drop the hypothesis.
What this graph is good for now
Relative-strength rotation and pair / spread mean reversion fit best. Narrow-industry PCA residuals can flag names stretched versus their industry factor, but you only get the latest residual z — history still has to be recomputed. Lead–lag and regime work as filters, not as the sole reason to enter.
Do not lean on the Corrada graph for intraday, volume, or VWAP strategies today.
What not to expect
- The graph will not hand you production parameters or position sizes
- Edges are stored when they can be computed, then filtered at query time; top edges are capped
- Prices are about two years of daily bars, not ticks
- PCA baskets are a narrow-industry snapshot, not a full factor model
- This is research and brainstorming, not investment advice
One line
Ask the graph “who is worth researching,” then ask the backtest “did this hypothesis survive out of sample.” If the sentence in the middle is vague, Corrada cannot help.