How Alphanume Research Tests Quantitative Trading Strategies

The actual workflow from a market mechanism to a frozen dataset, a hostile backtest, and a result worth publishing even when it fails.

Our editorial standard is to start a strategy study with a claim that could be wrong. A filing may predict dilution pressure, an exchange event may change short-horizon returns, or a volatility measure may identify an unusually expensive option. The target workflow turns that claim into a table with timestamps before interpreting a P and L curve.

That order matters. Most broken backtests do not fail because the researcher cannot calculate a return. They fail because today's clean database quietly rewrites what was knowable on the historical decision date, or because the final strategy remembers dozens of discarded versions.

This is the process we want readers to audit as of September 4, 2026. It is an editorial standard, not a claim that every older post contains every artifact below. Compare it with the publication's public description and inspect the resulting work in the research archive.

Start with a mechanism

A usable hypothesis explains who is acting, what constraint they face, when the information appears, and why a price effect might remain after costs. The mechanism does not need to be complicated. It needs to generate a prediction that can lose.

Suppose a company files a shelf registration. The vague story says dilution is bad. The testable version specifies the filing type, first public timestamp, eligible universe, entry delay, outcome window, financing context, and comparison group.

• Actor: issuer, trader, dealer, lender, index fund, regulator, or another identifiable participant.
• Constraint: capital need, forced execution, inventory risk, mandate, deadline, or information-processing delay.
• Clock: the first moment the relevant fact could enter the strategy.
• Prediction: a direction, distributional change, conditional probability, or relative-value spread.
• Failure condition: an outcome that would make us drop or revise the idea.

Our guide to mechanism-first trading edges expands this step. The point is to make the causal story earn its place before historical noise starts suggesting features.

Freeze the information set

Next we rebuild the decision as it existed then. Event time, ingestion time, effective date, and revision time can all differ. We preserve those clocks separately and join each feature as of the decision timestamp.

The Alphanume dataset catalog is designed around reusable event and market tables, while the API documentation exposes field definitions and filters. We still save the exact response used in a study because a reproducible result needs the snapshot as well as the endpoint.

Rows with uncertain clocks go into an exception ledger. Silently forcing them into the nearest session creates precision in the file and ambiguity in the result.

Build the boring baseline

The first run should be simple enough to audit without a notebook. We count events, dates, symbols, missing outcomes, repeated events, and exposure by cohort. Then we compare the proposed rule with a broad baseline and one mechanism-relevant control.

1. Define the universe. Use membership known on each historical date.
2. Attach outcomes. Keep unavailable and unresolved outcomes as nulls with reasons.
3. Choose a benchmark. Match market, sector, size, beta, or event context where the mechanism requires it.
4. Plot the distribution. Means can hide skew, fat tails, and a tiny set of extreme winners.
5. Break out time. Show years or regimes before combining the full history.

A strategy can look good because it is long beta in a bull market, short tiny companies during a crash, or concentrated in three dates. The baseline tells us which exposure actually did the work.

Make costs hurt

We separate signal quality from portfolio returns. A clean event effect may remain untradeable because the relevant names have wide spreads, unavailable borrow, tiny displayed size, violent gaps, or auction-only liquidity.

The execution model starts from a price available after the decision, then adds costs by the liquidity and holding-period assumptions the strategy would face. We report gross evidence first and a range of net cases after it.

• Use executable bid, ask, auction, or next-interval anchors instead of convenient bars.
• Model commissions, spread, slippage, financing, and borrow where applicable.
• Track no-fills and rejected orders as outcomes, not missing rows.
• Cap daily, sector, and ticker exposure before the cumulative curve exists.
• Compare equal-weight, volatility-scaled, and capacity-aware constructions only when each one answers a real portfolio question.

A result that survives an unfriendly cost range earns another test. A result that needs the reopening low, free borrow, and unlimited size has already answered the question.

Count every attempt

Search is part of the risk. Our piece on quant research at industrial scale explains why an AI agent can become an extremely efficient p-hacker when the reward only cares about the best submitted Sharpe.

Human research has the same problem at a slower speed. Our target workflow keeps a trial ledger containing every meaningful universe, feature, threshold, horizon, and portfolio variation. Any final result should be interpreted in light of how much searching produced it.

We prefer neighboring parameters that behave similarly, repeated evidence across subperiods, and a simpler construction when complexity adds little. A sharp optimum surrounded by failure gets treated as a warning.

Try to break it

Once the baseline works, the research turns hostile. We move the entry later, worsen costs, remove the most profitable dates, rebalance the universe, and test other regimes. We also search for mundane data explanations before telling a behavioral story.

• Timing probe: Delay the signal and use a less favorable executable anchor.
• Influence probe: Remove the largest winners and the busiest ticker-days.
• Universe probe: Exclude the smallest or least liquid securities.
• Parameter probe: Move each threshold to nearby values.
• Regime probe: Split calm, volatile, rising, and falling markets where relevant.
• Data probe: Rebuild a sample manually from primary filings or exchange messages.

We do not expect every strategy to pass every probe. We want to know exactly which assumption carries the result and whether that assumption has a defensible mechanism.

The public-data reconstruction in our spoofing case study is a good example of using raw records to understand the event before generalizing it into a trading claim.

Publish the misses

Research credibility comes from what happens after the attractive chart. In our strategy scorecard, we revisited live and published work with drawdowns and dead trades included. That follow-up is part of the method, not a marketing appendix.

A null result can close a research branch, refine a dataset, or expose a condition that future tests need to control. We keep those outcomes because deleting them makes the next idea look more original and the historical success rate look much higher than it was.

1. State the original rule. Readers should know what was frozen before the outcome.
2. Show the denominator. Include all eligible events and unresolved rows.
3. Report the ugly period. Drawdowns and regime failures explain the strategy better than a full-sample average.
4. Separate evidence from deployment. A historical effect does not prove that every reader could capture it.
5. Schedule the revisit. Recheck a published strategy after new data arrives.

The handoff

A finished Alphanume study should hand another researcher enough material to rebuild the sample. That means raw source snapshots, schema notes, query parameters, event construction, exclusion reasons, trial history, cost assumptions, and the exact tables behind the claims.

The prose then does a smaller job: explain why the question mattered, what we did, where the result changed, and what remains uncertain. A reader should leave with a method to reuse even when the specific trade is irrelevant to them.

• Can another person recreate the eligible cohort?
• Can they identify every field known after the decision time?
• Can they see every material research fork?
• Can they rerun gross and net results under different costs?
• Can they find the weak regime and the largest contributors?
• Can they tell which statements are measured facts and which are interpretation?

That is the standard we are aiming for. We will still get things wrong, and the useful response is to make the correction traceable and rerun the test.

If this is how you like to read quant work, inspect several studies in the archive linked above. Subscribe only after deciding that the process, including the misses, is worth following.