Methodology
How the Farlens composite score is built, how it is tested, and what the tests actually said — published in full, because a score you can't interrogate is an opinion with a number attached.
The composite score
Every name Farlens covers — equity or crypto — gets a composite score built from twelve weighted measurements across three families. Every individual weight is published on the signal stack page; the summary is:
- Price and money flow (54%) — trend, money flow index, on-balance volume, momentum, accumulation/distribution and volume surge. Deliberately weighted toward measurements that track money rather than price alone.
- Positioning and flow (34%) — crypto derivatives positioning (open interest, funding rates, liquidations), premium or discount to the 200-week moving average, sector rotation against the sector ETF, and FINRA off-exchange short volume.
- The world outside the chart (12%) — geopolitical event exposure mapped to affected sectors, and an engagement-weighted social narrative signal scored against each name's own trailing baseline.
Two facts worth stating plainly, because they cut against how this kind of product is usually marketed. The geopolitical overlay carries 6.9% of the composite — it is the most distinctive thing we measure, and it was never the whole score. And crypto is not a bolt-on: the two crypto-specific components together carry 19%, nearly three times the geopolitical weight.
Credit conditions are tracked as a separate market-wide backdrop rather than a per-name component — see credit spreads, explained. Components are weighted and combined into a single bounded score, and two design rules matter more than the weights themselves:
- A missing signal is missing, not neutral. When a component can't be computed, it is excluded from the composite rather than entered as zero — a zero would silently dilute every other signal and read as "no opinion" when the truth is "no data".
- Scores are computed fresh at read time, never served from yesterday's log — a cached score is a one-day lag wearing a current timestamp.
How it was tested
The composite was evaluated on roughly five years of daily observations — 13,872 ticker-days across the covered universe — using walk-forward validation: the score is fitted on one span of history and judged only on data it has never seen, rolled forward across three folds. Both the in-sample and out-of-sample legs are annualised before comparison, so a shorter test window can't masquerade as weaker performance.
Walk-forward matters because the obvious alternative — a single train/test split — flatters. We know because we ran both:
| Signal | Single 70/30 split | 3-fold walk-forward | Grade |
|---|---|---|---|
| Momentum-only (RSI) | 1.31 — looked robust | −0.24 retained; negative in one fold | OVERFITTED |
| Farlens composite | — | ~0.35 retained; positive in every fold | WEAK — but consistent |
Read that table honestly: the single-split test called a momentum-only signal robust, and the harder test showed it losing money out-of-sample. The composite grades WEAK on our own scale — it does not grade robust, and we won't claim otherwise. What it does do is stay positive in every out-of-sample fold, which is the property we actually care about: consistency under data the model has never seen.
What the score is not
- It is not a recommendation. Farlens publishes no buy/sell calls, no price targets, no entry or exit levels — on any surface, ever.
- It is not a prediction of any single outcome. It is a summary of present conditions across twelve measured inputs, with the components visible so you can disagree with any of them.
- It is never presented one-sided. Every read carries the bull case and the bear case together.
Historical analogues
For recurring event types, Farlens reports what actually followed past occurrences — "this condition has occurred N times since 2021; here is the distribution of what markets did next" — with sample sizes stated, because a median with n=4 is an anecdote wearing a statistic's clothes. Analogue counts grow as the event log accumulates; where history is thin, we say so.
Data sources and boundaries
Farlens builds on licensed geopolitical event data and public market data. The analysis layer — scoring, exposure mapping, analogues — is Farlens's own. Live readings are available to account holders; explanation, methodology, and history are public. We do not republish our upstream sources' raw feeds.