Building the search engine for recorded market microstructure
Replay answers one question: what happened here?
The research engine now answers the larger set. Where else did this happen? What was knowable at the time? What followed? How did matched periods compare with every other eligible bucket? Soon, how did one fixed population differ when a second setup-time condition was present, absent or false?
EdgeDepth records and replays full order flow across Binance USDT-M crypto and TradFi perps, then makes that record searchable directly, through the API, or from AI agents over MCP.
This is an honest view of what is live, what is in active rollout, what comes next, and what is still exploratory.
Available in the product today.
-
Live order-flow terminal
A browser-based terminal for the reconstructed DOM, time and sales, liquidation activity, VPIN, positioning, volume analysis, and market context.
-
30-day rolling replay
Pro users can replay the recorded market across the supported Binance USDT-M universe without having to record the session locally.
-
Exact seekable order-book replay
Seek directly to a recorded moment and inspect the reconstructed DOM, tape, and event context. Deep-seek DOM state has been manually verified across multiple pairs.
-
Permanent event archive
Important cascades, squeezes, pumps, crashes, macro events, outages, and volatility episodes remain available beyond the rolling replay window.
-
7-day free event replay
Authenticated free users can open eligible recent archived events with the normal recorded replay streams, including the order book. Older events require Pro unless an administrator marked the event public or free.
-
Guided lessons and Course Studio
Authors can turn recorded events into guided lessons with annotations, questions, and replay checkpoints.
-
Public showcase replays
Selected events can be opened without an account.
-
Market microstructure search engine
Search the complete eligible record across 660+ Binance USDT-M crypto and TradFi perps. The live 33-feature grammar covers order flow, price action, volatility, candles, funding, open interest, positioning and liquidations.
-
Research with receipts
Every scan reports matched and eligible counts, forward return plus favorable and adverse excursion distributions over 30 minutes, 1 hour, 4 hours and 24 hours, and the exact revision needed to reproduce the answer.
-
Historical and live moment reads
Read the frozen feature state at a recorded minute or inspect a just-happened bucket through the same grammar. Live reads are explicitly unpinned and provisional, then hand off to an ordinary reproducible history scan.
-
Direct, API and MCP research access
Run the deterministic engine in the browser, through the REST API, or from Claude, Cursor, Codex and other AI agents over MCP. API and MCP access also include matched-cohort comparison against every otherwise-eligible baseline bucket.
-
Reports, definitions and live alerts
Publish a pinned result, save the exact research definition, or arm it as a read-only live condition with in-app, Telegram and Discord delivery.
Built foundations that are not yet fully public across the production experience.
-
21 curated conditions
The /signals library joins curated conditions to pinned historical frequency, full forward outcome distributions and reproducibility keys. The pages and nightly receipts are live but remain outside search indexing while the current-bucket firing feed and refresh schedule finish production ratification.
-
Browser cohort comparison
Matched-versus-baseline comparison already runs through the engine, REST API and MCP. The browser comparison view remains behind its rollout gate until the production presentation is ratified.
-
Fixed-anchor stratified comparison
The backend release candidate is independently audited and ready for deployment. It freezes one population of exact occurrence anchors, classifies each anchor as split true, false or absent, and gives every closed group the same outcome summaries. Direct backend smoke, then separate public API, browser and MCP release gates still remain.
-
Search the whole event archive
Structured server-side archive search with normalized filters, archive-wide facets, stable pagination, shareable query URLs, deterministic match evidence, replay-access state, and selected-archive disclosures.
The next credibility and research-depth work, in intended order.
-
Hyperliquid and deeper verified history
Hyperliquid support is coming soon. The planned expansion targets up to two additional years of Binance Futures and one to two years of Hyperliquid history. Raw ticks become searchable research only after venue-specific integrity checks, point-in-time feature extraction, coverage manifests and deterministic backfill verification pass.
-
Robustness and concentration diagnostics
Show whether a historical result is spread across symbols and days or concentrated in a few episodes, with coverage and blocked stability checks. The recorded k-of-n count stays exact; EdgeDepth will not add a misleading independent-trials confidence band to dependent market events.
-
Design a parallel feature library v2
Design a new, opt-in feature version for CVD, order-flow imbalance, absorption, OI-price divergence and cross-market context. Version 1 remains frozen; cross-symbol leakage and replay parity must be proved before any public backfill.
-
Durable research receipts
Promote scan and condition receipts into a reproducibility-keyed backend store so useful findings survive process restarts and can support a public research library.
-
Archive verification and repair
Verify new archives before publication, preserve provenance, quarantine factual failures and classify older assets by what can be trusted or deterministically repaired.
Ask a market question, then inspect the evidence
Find every eligible BTCUSDT, NVDAUSDT and XAUUSDT period where the book was ask-heavy, open-interest velocity ranked in its top decile, top traders were at least 15 points more long than the crowd, and at least $5 million was liquidated in the trailing hour.
EdgeDepth translates that question into a validated, versioned definition, scans the complete eligible daily record, preserves what was knowable at the setup time, calculates later outcomes separately, and returns a reproducible set of historical windows.
Each result carries its source symbol and moment back into the same tick record, so the evidence can be inspected in replay.
This is historical research, not a signal and not a promise of future performance.
Directions under consideration, with no delivery promise.
-
Research-to-lesson workflows
Turn reproducible historical findings into replay-linked teaching material.
-
Cross-exchange research
Start by running one pinned definition separately on Binance Futures and Hyperliquid with venue-labeled counts, coverage and outcomes. Later paired studies can add explicit time alignment, source-venue anchors and venue-qualified features for leader-lag, basis and funding questions without silently pooling unlike contracts.
-
Mobile archive and lesson views
Focused archive browsing and learning on smaller screens.
-
Transparent calibrated scoring
A descriptive score is worth exploring only if it ships with visible reliability evidence and stays secondary to raw counts, denominators and outcome distributions.
The newest completed research and replay foundations.
-
Wave 3 market-stress and positioning features
Open-interest velocity, one-hour liquidation notional, top-trader and global long ratios, and top-trader-versus-crowd positioning skew are backfilled and active across the research store.
-
Four-horizon outcome distributions
Research results now report 30-minute, 1-hour, 4-hour and 24-hour returns plus maximum favorable and adverse excursion over every complete occurrence.
-
Matched cohort comparison
The engine, REST API and MCP can compare what followed a setup with what followed every otherwise-eligible bucket, preserving counts and the full distributions without inventing a lift score.
-
Read the market now
A live snapshot reads the newest market bucket through the same feature grammar, labels forming values provisional, and turns that observation into a reproducible historical question.
-
REST and agent-native research
The public API and read-only MCP server expose registry discovery, interpretation, scans, pagination, snapshots, base rates, commonality, reports and cohort comparison with canonical response bytes.
-
Search-to-replay evidence
Research occurrences carry their source symbol and moment back into the same full-depth tick record used by the terminal and replay system.
The market already happened. Make it useful.
EdgeDepth began as a way to replay a move you missed.
It is now a system for finding, verifying, comparing, and understanding recorded market behavior.