Flow · Flows are not trading intent
Research flow availability times, label errors, and incremental predictive power.
Observation
On-chain transfers, exchange net inflows, fund flows, or stablecoin changes may precede prices, but each measures different objects. A wallet transfer may be internal; reported fund data may arrive after trades. Temporal association alone is insufficient.
Mechanism
Genuine external capital inflows may bring future buying; collateral deposits may instead prepare shorts or maintain margin. Rising prices also attract capital. Distinguish informational content, reverse causality, and common drivers.
Hypothesis
Teaching hypothesis: after controlling for past returns and volatility, can external net inflows visible in the preceding period improve next-period return forecasts? Fix the primary metric, window, holding horizon, and data version beforehand. Without historical release vintages, current data supports exploration only, not a claim of historically tradable validation.
Data
Store event time, first visibility, retrieval time, source version, wallet/product labels, and revisions for every row. Check duplicates and internal transfers before treating outliers; do not delete observations for contradicting the expected direction. Report missing-label rates and asset coverage over time.
Signal
Normalize size using a predeclared method, such as dividing by the previous circulating supply. Denominators must not use future revisions. Activate a signal at the next executable time after publication. Stale or missing data means no signal; do not carry it forward until favorable results appear.
Backtest
Compare a “past price information” baseline with “baseline plus flow” on the same test set, evaluating incremental returns and errors. Add release delays, label noise, and shrinking coverage variants. All variants share costs and position limits.
Falsification
If the edge disappears after delaying to actual publication, the original result may be look-ahead. If removing internal transfers or controlling for price trends erases it, withdraw the flow mechanism. Use a negative control: if information unrelated to demand for the target asset also appears significant, investigate shared time trends.
Exercise and project acceptance
Build an event ledger with four times: “trade occurred,” “data published,” “system retrieved it,” and “order permitted.” Deliberately include one label revision and missing data. Submit visibility-filtered signals and an exclusion record. A reviewer must reconstruct what was known then, not merely recalculate today's tidy dataset.
Connect the data plan to Advanced C · Alternative Data and allocation to the capstone handbook.