Calibration & Backtest Limitations

This is the most important page on the site. graphGURU.ai is built to be honest about what it can and cannot do, so read this before you rely on anything you see in the app.

Not financial advice. This is an information tool with real, documented limitations. Nothing here is a recommendation to buy, sell, or hold any asset. Every trading decision — and every consequence of it — is yours alone.

The edge is concentrated in the past, not the present

Our signals were calibrated against roughly six years of Binance candle history. When we split that window, the measurable edge is heavily concentrated in the 2020–2023 period. On the most recent stretch — the last ~2.4 years — the same strategies land close to break-even at 1× leverage, before fees. In other words, the strong historical numbers are largely a product of an earlier market regime that may not repeat.

Some timeframes are essentially a coin-flip

Timeframe matters enormously. Over the full six-year test, raw signal confidence has essentially no ranking power on direction at the 4-hour timeframe — directional calls there land around 49.6%, which is a coin-flip. The daily (1d) timeframe carries meaningfully more signal. We do not hide this behind an inflated confidence number: short intraday timeframes are the noisiest part of the system, and you should treat them accordingly.

Why confidence tops out near 54

The "confidence" figure is deliberately calibrated so the number approximates the real historical hit-rate. It sits on roughly a 49–54 scale on purpose — nothing scores 70+. Earlier "70–92%" style claims, when actually measured, only delivered about 52–59%. The separate signal strength value (a raw ~20–92 score) reflects how many indicators agree, not a probability. Do not read a high signal-strength number as a high chance of being right.

Proxy calibration and thin history

Not every timeframe has its own calibration curve. Several timeframes use proxy curves borrowed from a related timeframe, which is an approximation, not a direct fit.
Newly-listed coins (for example XAUT) have limited price history, so their calibration leans on a shared/global map and has less data behind it.
The 15-day and 30-day timeframes are synthetic — built by grouping daily candles, not native exchange intervals — so they have far fewer data points and some indicators are incomplete.

Backtests can overfit, and the past is not the future

Any strategy tuned on historical data risks overfitting — fitting the noise of one dataset rather than a durable effect. We try hard to guard against this (out-of-sample validation, refusing to ship signals from short windows), but the risk never goes to zero. Live results routinely differ from historical replays, and past performance does not predict future results.

Fees, slippage, and leverage risk

Backtest figures are gross of real-world frictions. Actual trading incurs fees and slippage that erode — and on marginal strategies can erase — any edge. If you use leverage, an adverse move of just 1/N of your position (for N× leverage) can liquidate you for a total loss. Crypto is volatile and can move against you faster than any stop can protect you.

You bear all risk. graphGURU.ai provides automated analysis of public market data for information only. It is not financial, investment, or trading advice, and it does not know your circumstances. You are solely responsible for your own decisions and their outcomes. Never risk money you cannot afford to lose, and do your own research.