What I built
I built the leakage-safe signal-to-execution path with frozen assumptions, auditable scenarios, and checksummed artifacts.
Research question
When do order-flow imbalance, liquidity, and limit-order-book conditions predict short-horizon price movement—and which claims survive data-quality and execution gates?
Predictability, executability, and profitability are separate questions. The project makes event ordering, temporal leakage, gaps, fees, latency, uncertain fills, adverse selection, and inventory risk explicit before a claim can be promoted.
Execution stress test
Research reference only. One deterministic four-hour market-order simulation; not live trading. The two adjacent pseudo-heldout phases are neither prospectively untouched nor cross-day replication. The 144 scenarios overlap and must not be summed.
v2.0 Fee & Queue-Position Re-Pricing of the 144 Market-Order Scenarios (exploratory)
| Execution Tier | Fee (bp) | Adverse Selection (bp) | Effective Hurdle (bp) | Gross > 0 | Net > 0 (of 144) | Best Net Edge (bp) | Median Net Edge (bp) |
|---|---|---|---|---|---|---|---|
| Retail Taker (Frozen Baseline) | 4.00 | 0.00 | 4.00 | 110 | 0 | -1.6148 | -3.8555 |
| VIP-9 Taker | 1.50 | 0.00 | 1.50 | 110 | 9 | +0.8852 | -1.3555 |
| Maker 0 bp (Tail-of-Queue) | 0.00 | 1.85 | 1.85 | 110 | 9 | +0.5352 | -1.7055 |
| Maker 0 bp (Median-Queue) | 0.00 | 1.15 | 1.15 | 110 | 9 | +1.2352 | -1.0055 |
| Maker 0 bp (Front-of-Queue) | 0.00 | 0.45 | 0.45 | 110 | 19 | +1.9352 | -0.3055 |
| Maker Rebate -0.50 bp (Median-Queue) | -0.50 | 1.15 | 0.65 | 110 | 11 | +1.7352 | -0.5055 |
How to read this table. Each row subtracts a different hurdle (fee plus an assumed adverse-selection cost) from the gross edge of the same 144 deterministic market-order scenarios; no passive order was simulated, and the assumed fill probabilities (0.42–0.74) are not applied to any figure. “Best net edge” is the maximum over 144 overlapping scenarios minus the hurdle, with no multiple-testing adjustment. That maximum (2.3852 bp gross) comes from nine latency variants of one ETHUSDT primary-phase 5-second scenario with 10 fills, so +1.9352 bp is a selected single-session observation, not an achievable maker edge.
Earlier trade-only protocol terminal
This separate, earlier M8 study remains a valid NOT_RUN terminal. It should not be confused with the later nonconfirmatory continuous L2 pilot summarized above.
| Stage | Rows inspected | Warnings | Outcome |
|---|---|---|---|
| BTCUSDT training | 2,071,461 | 0 | Normalization passed. |
| ETHUSDT training | 987,297 | 53 | The predeclared zero-warning gate failed. |
| Model selection | 0 | — | Not started. |
| Held-out evaluation | 0 | — | Dates were not opened. |
The valid terminal is INSUFFICIENT_DATA. It is evidence that the protocol stopped correctly—not evidence for or against a market hypothesis. Execution and P&L remain NOT_RUN.
Capabilities demonstrated
- Preserve exchange and receipt ordering with explicit continuity epochs.
- Construct features only from information observable at the decision point; labels begin strictly afterward.
- Use purged, embargoed, time-ordered model selection and prediction-only evaluation.
- Carry predictive diagnostics through fees, latency, fills, adverse selection, inventory, and liquidation assumptions.
- Freeze checksummed run bundles and generate reports from artifacts instead of copying preferred metrics into prose.
Public data boundary
The public GitHub and Dataset packages contain source code, configuration, tests, protocols, and documentation. They do not redistribute exchange observations, normalized or derived events, fitted states, ingestion authorities, or generated run bundles. The website and Space expose one machine-generated, low-dimensional 144-scenario summary with checksummed provenance; it contains no raw events, ledgers, models, or local paths. Users obtain public market data independently under provider terms.
Reproduce and inspect
Run make setup, make check, and make reproduce-sample for the deterministic synthetic vertical slice. Synthetic output verifies software behavior only. The public exploratory summary was generated from a checksum-verified local bundle, not hand-copied into this page.
Limitations
- The current result uses one continuous four-hour capture and deterministic market-order assumptions.
- Its adjacent evaluation phases reuse overlapping source data; there is no prospectively untouched or cross-day confirmation.
- The 144 latency, horizon, symbol, and phase scenarios are alternatives, not positions in a portfolio.
- No result is a live-trading, significance, capacity, profitability, annualized-return, or investment claim.