The Research Question: Do mid-cap US equities ($4B to $20B market cap) carry intraday microstructure patterns that survive realistic trading costs? Mid-caps are liquid enough to trade, but less efficient and less crowded by high-frequency traders than mega-caps, so institutional footprints can still show up in order flow. In January 2026 I set out to answer that with a disciplined research process: define the tests in advance, try hard to break every result, and treat "no edge" as a valid answer.

What I Tested

17 research objectives across order flow, liquidity and intraday behaviour: an original set of eleven, plus six added once the first results showed where the most promising directions were. Each had its own pre-registered hypothesis and acceptance criteria. A few examples:

• Quote imbalance: when bids heavily outweigh offers and the spread is tight, is the next 60-second return positive?
• Liquidity evaporation: does rapid spread widening predict a volatility spike in the next minute?
• Off-exchange prints: does heavy off-exchange volume during consolidation predict a breakout?
• Volatility compression: when Bollinger Band width falls to the bottom of its six-month range, do the breakouts that follow sustain?
• Sector decoupling: when a sector ETF drops but a stock holds up, does the stock outperform over the next 15 minutes?

Research Standards

Every objective ran under the same non-negotiable standards:

• Pre-registration: each objective's hypothesis, method and acceptance criteria were written down before it was run (for example: out-of-sample return above zero net of costs, p < 0.01, and every negative control collapsing), and the total number of tests was reported.
• No leakage: every feature had to be computable from information available at decision time. Any use of future data was a hard failure.
• Negative controls: promising results had to face three fraud checks: shifting the signal forward one bar, randomizing entry timestamps within the day, and permuting returns. A real signal should collapse under all three; if it didn't, leakage was assumed.
• Multiple-testing control: global Bonferroni correction across all tests, with bootstrap confidence intervals and permutation tests.
• Walk-forward validation: fixed training, validation and test splits, with purged cross-validation and an embargo for overlapping labels.
• Execution realism: commissions and slippage tied to spread and volatility. If a mild cost increase flipped the result, the edge was declared not real.

The Multi-Agent Pipeline

I designed the research program, its standards and the agent governance, then ran the lab as a small team of AI agents:

• Research agents: up to five Claude Code agents running concurrently, each owning one objective under a "Senior Quantitative Researcher" brief whose mandate was to protect capital by invalidating weak ideas.
• Independent auditor: a separate ChatGPT Codex agent with a "hostile but fair" brief reviewed each agent's work for protocol compliance, leakage and statistical rigour. It caught real defects, including a daylight saving time bug and lookahead leakage in a quantile-binning step.
• Standard pipeline per objective: data loader (Polygon.io API, caching, validation, corporate actions) → features → signals → cost-aware simulator → statistics → automated report.
• Data and compute: ~27,000 Polygon.io API calls and ~1.35 billion trade and quote records, processed on a 12-vCPU cloud server over about three weeks, with raw data kept immutable and feature datasets versioned.

Results

• 10 of the 17 objectives reached a verdict by the January checkpoint. The rest were still in progress or blocked on data when the lab paused.
• 7 were rejected with documented reasons: no edge, not statistically significant, or not surviving realistic costs.
• One candidate edge: volatility compression. Breakouts after extreme Bollinger Band compression showed statistically significant trend continuation, consistent with volatility clustering.
• One risk filter: rapid spread widening reliably predicts volatility spikes. It isn't a standalone entry signal, but it is useful for position sizing and risk control.
• One inconclusive: the institutional retrace objective was marked for another iteration rather than forced to a verdict.
• The most instructive kill: quote imbalance showed an edge in raw analysis that disappeared entirely once realistic transaction costs were applied.

A high kill rate is the point. A lab that finds edge everywhere is lying to itself. Most hypotheses should fail, and here they did, each for a documented reason.

Philosophy and Status

The lab follows a "grey box" model: AI agents act as the signal engine (scanning, testing and reporting), while a human keeps final control of execution and risk. The program is paused, not abandoned. Two follow-on tracks are specified: formalizing a discretionary small-cap intraday strategy into a causal, testable backtest, and a scoring engine that ranks stocks by asymmetric post-earnings reaction potential.