Research notes
Short, self-contained studies on single questions. None are published yet - this page documents the format and the standards each has to meet first.
There are no research notes on this page because none have been completed to the standard below. Publishing plausible-looking numbers ahead of the work would be straightforward and would defeat the purpose of the domain. The queue and the format are public so that what eventually appears can be checked against what was promised.
Format
Every note follows the same six sections, in this order.
| Section | Contains |
|---|---|
| Question | One question, narrow enough to answer. Not a topic. |
| Method | Exactly what was computed, including every parameter chosen and why. |
| Dataset | Universe, date range, observation count, and what was excluded. |
| Observation | What the data showed, stated without interpretation. |
| Limitations | What could have produced this result other than the effect. Written before the conclusion, deliberately. |
| Conclusion | What the observation supports — usually less than it first appears to. |
Limitations precede the conclusion because writing them afterwards turns them into caveats attached to a result already decided on. Writing them first changes what the conclusion is allowed to say.
Standards a note must meet
- Point-in-time data only. Any analysis of whether one series leads another must use values as they were available at the time. Backfilled or revised data produces look-ahead bias, and look-ahead bias produces results that look excellent and are worthless.
- Survivorship stated explicitly. Delisted and abandoned assets are missing from most datasets. Any study of long-horizon outcomes must say what happened to assets that disappeared, or acknowledge that it cannot.
- Sample size reported, not implied. Number of assets, number of periods, number of events. A finding drawn from twelve events is described as drawn from twelve events.
- Base rates included. "Social volume rose before 70 % of large moves" is uninformative without knowing how often it rose before everything else.
- Costs included where relevant. Any note touching implementable strategies accounts for spread, slippage and the fact that the assets with the strongest signals are usually the least liquid.
- Reproducible. Enough detail that someone with the same data reaches the same number. If the data cannot be shared, that is stated.
- Negative results published. A study finding no relationship is published as readily as one finding a relationship. Selecting only positive results is how a research programme becomes a marketing programme.
On backtests specifically
Backtests are conspicuously absent from the queue below. That is deliberate. Moonboard's percentile-ranked metrics are computed against the universe as it exists at each refresh, and reconstructing historical universes correctly — including assets that have since disappeared, with their data as it stood then — is a substantially harder data problem than running the strategy logic.
Until that reconstruction exists and has been checked, any backtest of these metrics would be measuring the reconstruction rather than the strategy. No backtest results will be published before the historical data supporting them is documented here.
Open questions in the queue
Listed with the reason each is difficult, since that is usually the more useful information.
- Does social momentum lead price, or follow it?The premise behind collecting social data at all. Requires strictly point-in-time data and a definition of "lead" that survives the fact that both series respond to the same news.
- How does sentiment behave during high-volatility periods?Tests the suspected bias that classifiers read crisis-era community language as positive. Needs human-labelled text from drawdown windows as a comparison.
- Is social activity elevated before large price moves?Distinct from the first question: about tail events rather than average relationships. Base rates matter — most social spikes precede nothing.
- How stable are percentile rankings between refreshes?Directly quantifies how much of the movement in the ranked metrics comes from the universe rather than from the asset. Requires no external data — measurable today.
- How much does sector classification choice change measured rotation?Compares the same window under several schemes. Bounds how much of a rotation finding is an artefact of the taxonomy.
- Does developer activity persistence correlate with project survival?Blocked on survivorship bias: dead projects are missing from most repository datasets, and they are the observations that carry the answer.
- How concentrated is measured social volume by source?If a small number of accounts produce most posts for an asset, social volume is measuring those accounts rather than a community.
- Does the Moonboard Score volume component discriminate at all in practice?The cap binds above 0.3 % turnover. A distribution check would establish what share of the universe receives full marks, and whether the component is effectively constant.
Related
- Research areasWhere these questions come from
- MethodologyWhat is being studied
- Percentile rankingWhy historical reconstruction is hard here
- Data sourcesWhat data exists to study