MicroAlpha Execution + Alpha Research Lab

A research platform for limit-order-book execution, short-horizon alpha evaluation, and robustness under market shift, with explicit transaction-cost modeling and results reported in net basis points after costs.

This repository studies three related but distinct questions:

1. Can order-book information improve execution quality relative to standard schedules such as TWAP and VWAP?
2. Can microstructure-aware signals remain net positive after costs on real historical data?
3. Can a strategy remain useful when the target market differs from the training market?

The project is organized as a four-phase benchmark:

1. Phase 1 — Execution research
   Event-driven benchmark on real and synthetic LOB data with implementation-shortfall metrics in basis points.

2. Phase 2 — Alpha research on a real-calibrated synthetic universe
   Cost-aware market-neutral benchmark with multiple cross-sectional baselines.

3. Phase 3 — Fully real walk-forward alpha benchmark
   Anchored day-by-day evaluation using only real AAPL LOB days.

4. Phase 4 — Real multi-asset / cross-venue robustness
   AAPL walk-forward plus Binance transfer test with a domain-shift fallback mechanism.

The goal is not to claim production live trading.
The goal is to make the contribution of microstructure-aware methods legible under costs, turnover, and regime shift.

---

Why this repository is useful

Many trading side projects focus on prediction alone, rely on toy backtests, or omit realistic cost accounting.

This repository instead emphasizes:

• real microstructure data
• walk-forward evaluation
• spread-based transaction costs
• clean, comparable baselines
• net bps, turnover, hit rate, and episode-level stability
• hybrid ML + microstructure decision rules
• robustness under transfer shift

---

Key results

Phase 4 — combined real multi-asset benchmark

On the combined benchmark
(AAPL / XNAS walk-forward + BNBUSDT / BINANCE external transfer test):

• CAMP-R: 0.1711 mean net bps / step
• RAMP-R: 0.1643
• online_regime: 0.1431
• ridge_static: 0.0798

Headline summary:

• 29,194.6 total net bps
• 9 positive episodes out of 10
• episode t-stat ≈ 3.70

Phase 3 — fully real walk-forward benchmark

On the default fully real anchored walk-forward evaluation:

• RAMP-R: 0.3148 mean net bps / step
• online_regime: 0.2582
• ridge_static: 0.1446
• micro_fixed: -0.0284
• momentum: -0.1140
• signed_flow: -0.1287

RAMP-R also delivers:

• 26,528.2 total net bps
• 7 positive days out of 9

Phase 1 — execution benchmark

Phase 1 compares:

• TWAP
• VWAP
• Liquidity-only
• Alpha-only
• MALP

using implementation shortfall in bps as the main execution metric.

---

Project structure

The repository is best read as a progression:

execution → controlled alpha sandbox → real historical alpha → robustness under shift

This separation is deliberate.

• Phase 1 answers an execution question: if a parent order is already known, how should it be split into child orders?
• Phases 3 and 4 answer alpha questions: should the strategy take directional exposure, and how should it control model trust under drift?
• Phase 2 provides an intermediate sandbox between execution and fully real alpha evaluation.

Because these are different problems, they use different baselines.

• In Phase 1, the relevant baselines are TWAP and VWAP, because the problem is execution.
• In Phases 3 and 4, the relevant baselines are directional rules such as momentum, signed flow, micro_fixed, ridge_static, and online_regime.

---

Main methods

Phase 1 — Execution research

A clean event-driven execution benchmark on real and synthetic LOB data.

Included policies:

• TWAP
• VWAP
• Liquidity-only
• Alpha-only
• MALP = Microstructure Alpha + Liquidity Policy

Why it matters

This phase answers the execution question:

can a microstructure-aware policy reduce implementation shortfall relative to naive schedules?

MALP combines two ingredients:

• local liquidity quality, which tells us where the book looks easier to access;
• execution alpha / urgency, which tells us when waiting may become more expensive.

This phase establishes that order-book information is useful even before any directional trading problem is introduced.

---

Phase 2 — Alpha research on a real-calibrated synthetic universe

Phase 2 adds:

• a synthetic multi-symbol universe calibrated from real AAPL order-book snapshots;
• cross-sectional market-neutral backtests with explicit turnover costs;
• RAMP = Regime-Adaptive Multi-horizon Portfolio.

Baselines:

• uniform
• imbalance
• momentum
• ridge5
• RAMP (ours)

Why it matters

This phase bridges pure execution research and alpha research by asking:

can short-horizon microstructure structure be turned into cost-aware cross-sectional alpha?

Phase 2 is intentionally not the strongest evidence in the project.
Its role is to provide a controlled research sandbox in which multi-horizon prediction, regime adaptation, and cost-aware allocation can be tested before moving to fully real historical evaluation.

---

Phase 3 — Fully real walk-forward alpha benchmark

Phase 3 removes the synthetic test universe and works only with real AAPL LOB days bundled in the repository.

It adds:

• anchored walk-forward evaluation
• a same-day warmup regime detector
• a ridge model fit only on prior real days
• sparse position-taking with explicit spread-based cost accounting
• RAMP-R = Regime-Adaptive Microstructure Portfolio on Real data

Baselines:

• momentum
• signed_flow
• micro_fixed
• ridge_static
• online_regime
• RAMP-R (ours)

RAMP-R in one paragraph

RAMP-R combines three ingredients:

1. a microstructure signal based on microprice displacement / imbalance;
2. a ridge predictor trained only on prior real days;
3. an online regime switch estimated from the first part of the current day.

If the warmup segment suggests that the usual imbalance relationship is informative enough, the model trusts the current-day sign. Otherwise, it falls back to a more stable prior-day logic. The final score is thresholded into sparse positions, held for a fixed horizon, and evaluated after spread-based turnover costs.

Why it matters

This is the first phase that answers the central alpha-research question:

can the strategy remain net positive on fully real out-of-sample days once costs are accounted for?

---

Phase 4 — Real multi-asset / cross-venue robustness benchmark

Phase 4 keeps the fully real AAPL walk-forward benchmark and adds a second real market:

• AAPL / XNAS historical LOB walk-forward
• BNBUSDT / BINANCE as an external transfer test

It adds:

• a real cross-asset / cross-venue net-bps benchmark
• CAMP-R = Cross-Asset Adaptive Microstructure Portfolio on Real data
• a simple domain-shift gate deciding when to trust the hybrid ML layer and when to fall back to safer microstructure logic

Why it matters

This phase tests whether the signal logic is merely in-domain fit, or whether it can degrade gracefully under transfer.

CAMP-R does not simply ask whether one model is strongest on average.
It asks a more realistic question:

when the current market looks different from the training market, should the same model be trusted to the same extent?

That is the core robustness contribution of the project.

---

Quickstart

Create a virtual environment and install dependencies:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .

Run the four main benchmarks:

python scripts/run_benchmark.py
python scripts/run_alpha_research_phase2.py
python scripts/run_alpha_research_phase3.py
python scripts/run_alpha_research_phase4.py

Launch the dashboard:

streamlit run app/demo_dashboard.py

---

Repository layout

.
├── app/
│   ├── demo_dashboard.py
│   ├── common.py
│   └── pages/
├── data/
│   ├── optiver/
│   └── binance/
├── docs/
│   ├── method_note.md
│   ├── alpha_research_method.md
│   ├── alpha_phase3_method.md
│   └── alpha_phase4_method.md
├── reports/
│   ├── benchmark_summary.md
│   ├── alpha_phase2_summary.md
│   ├── alpha_phase3_summary.md
│   ├── alpha_phase4_summary.md
│   ├── *.png
│   ├── *.csv
│   └── *.json
├── scripts/
│   ├── run_benchmark.py
│   ├── run_alpha_research_phase2.py
│   ├── run_alpha_research_phase3.py
│   ├── run_alpha_research_phase4.py
│   └── build_phase4_episode_summary.py
├── src/
│   └── microalpha_exec_lab/
├── pyproject.toml
├── requirements.txt
└── README.md

---

Dashboard

The Streamlit dashboard is organized to mirror the project logic.

Suggested reading order:

1. Overview
2. Phase 1 — Execution
3. Phase 3 — Real Alpha
4. Phase 4 — Robustness
5. Methods & Limitations
6. Phase 2 — Appendix

The dashboard is designed to clarify:

• what each phase studies
• which baselines are relevant in each phase
• how the reported metrics should be interpreted
• what the project models explicitly and what it does not

---

Scope and limitations

This repository explicitly models:

• spread and displayed top-of-book liquidity
• level-1 / level-2 imbalance
• signed trade flow
• spread-based transaction costs
• walk-forward retraining
• warmup adaptation
• domain-shift-aware fallback logic

It does not explicitly model:

• exact queue position
• passive fill probability
• detailed queue cancellation dynamics
• latency-sensitive exchange mechanics
• full production-grade order management

The correct interpretation is therefore:

a research-grade microstructure benchmark and alpha-validation platform, not a full exchange simulator

---

Why the benchmark design is coherent

The benchmark is designed so that each phase answers one main question with the appropriate comparators.

• Phase 1 compares execution policies against TWAP and VWAP, because the task is parent-order scheduling.
• Phase 2 tests cross-sectional alpha construction in a controlled sandbox.
• Phase 3 evaluates directional alpha on fully real historical data in a strict walk-forward setting.
• Phase 4 studies robustness under cross-market shift and adaptive model trust.

This separation is important because it prevents execution, prediction, and robustness from being mixed into a single uninterpretable benchmark.

---

Core quantities used throughout the project

The repository relies on standard market-microstructure quantities, including:

• midprice
• spread
• level-1 / level-2 imbalance
• microprice
• signed trade flow
• rolling returns
• turnover
• net bps after costs

In the execution phase, the main metric is implementation shortfall in basis points.
In the alpha phases, the main metric is mean net bps / step, together with cumulative net bps, turnover, hit rate, positive episodes, and episode-level stability.

---

Reproducibility

The repository is organized so that the main results can be regenerated from the benchmark scripts and inspected through the dashboard.

The intended workflow is:

1. run the benchmark scripts,
2. generate summary files and figures in reports/,
3. inspect phase-by-phase behavior in the dashboard.

If you modify the data splits, cost settings, or feature definitions, results will change accordingly.

---

Suggested entry points

If you only inspect a few files first, start with:

• app/demo_dashboard.py
• app/pages/1_Phase_1_Execution.py
• app/pages/2_Phase_3_Real_Alpha.py
• app/pages/3_Phase_4_Robustness.py
• docs/method_note.md

If you want to run only one benchmark first, start with:

python scripts/run_alpha_research_phase4.py

and then open:

streamlit run app/demo_dashboard.py

---

Summary

The central contribution of this repository is not a claim of perfect market simulation.

The contribution is a structured research framework that separates:

• execution quality
• directional alpha
• robustness under drift and transfer shift

In that sense, the project is best read as:

execute better → experiment safely → validate on real data → control model risk under shift