Backtest (simulated)
What gitforage's rules would have brought home on 26 past Sundays, and how often those finds later turned up on GitHub Trending. This is a replay on past data, not a record of what any portal showed.
Replayed Sundays to , outcomes read through , rules from gitforage at 79a3381.
What the replay found
gitforage ranks by fit to a reader's interests, not by virality, so this is not a test it was built to win. Still, the field notes are worth keeping. Across the full topic catalog, the finds that did go on to trend were shown a median of 54 days before their first Trending day (the middle half fell between 34 and 74 days).
The other side of the ledger matters as much. The replay showed only 85 of the 1,025 in-topic repos that reached Trending for the first time (8.3%), and 45 of 646 (7.0%) on the demo's five default communities. Most finds never trended at all: 67.9% of those shown at least eight weeks before the end never reached Trending or the Hacker News proxy. For a guide that picks what fits, that is the expected shape. Most good tools for a niche never trend.
| Run | Finds shown | Trending first-timers shown earlier | Days earlier: median (middle half) | Shown, never trended |
|---|---|---|---|---|
| Demo default (5 communities) | 1,033 (41 a week) | 45 of 646 (7.0%) | 55 (40 to 75) | 66.7% |
| Full topic catalog (12 groups) | 2,239 (90 a week) | 85 of 1,025 (8.3%) | 54 (34 to 74) | 67.9% |
| Demo digest (17 topics, top 20 a week) | 500 (20 a week) | 32 of 495 (6.5%) | 53 (24 to 76) | 40% |
A check with topics ignored
The same weekly budget, spent on raw star velocity across all of GitHub with no regard for fit, caught slightly more future Trending repos and showed fewer that never trended. Topic fit does not beat raw velocity at predicting Trending. It is not meant to: velocity alone would fill a field guide with whatever is loud that week.
| Run | Finds shown | Trending first-timers shown earlier | Days earlier: median (middle half) | Shown, never trended |
|---|---|---|---|---|
| Demo default, velocity only | 1,033 (41 a week) | 49 of 646 (7.6%) | 55 (13 to 75) | 43.9% |
| Full catalog, velocity only | 2,179 (87 a week) | 101 of 1,025 (9.9%) | 56 (17 to 76) | 52.1% |
On Hacker News (stories linking the repo with 100 or more points, standing in for the front page), the replay caught 8 of 154 in-topic first-timers early on the full catalog and 5 of 92 on the demo, a median of 30 days ahead.
The longest leads
From the full catalog run. A long lead may be a repo that grew slowly and then trended for an unrelated reason; these are reported as they are.
| Repo | Shown (simulated) | First on Trending | Days earlier |
|---|---|---|---|
| floci-io/floci-az | May 31, 2026 | Nov 3, 2026 | 156 |
| Mailu/Mailu | May 31, 2026 | Oct 24, 2026 | 146 |
| pnnbao97/VieNeu-TTS | Jul 26, 2026 | Nov 15, 2026 | 112 |
| vinceAmstoutz/symfony-security-auditor | Jun 21, 2026 | Oct 11, 2026 | 112 |
| lidge-jun/opencodex | Jul 26, 2026 | Nov 14, 2026 | 111 |
Caveats about the data
- The public star record thinned out. GH Archive, the public record of GitHub events, fell from about 600,000 star events a week in late 2025 to 10,000 to 60,000 a week from July 2026, roughly a tenfold drop. Over the window it holds about 2.5% of real stars.
- Rank order still holds. Checked against true star counts, the sampled stars keep their order (a median rank correlation of 0.82), so who is growing faster is still visible. Small projects' growth, and niche patches such as HL7 and FHIR, cannot be judged fairly this way.
- Topics and descriptions are today's. A repo that added a topic after it trended gets credit for a fit it may not have had then. The velocity-only runs are the check on that.
- Some rules could not be replayed: release recency, README checks, kinship with the demo's followed projects, and visitor feedback.
- Trending archives are kept by volunteers and miss some days, so a few Trending dates may be late.
- The first replayed week is left out, because it lists everything already big when the replay began.
The live record
From the public demo keeps a live early-finds record of what it actually shows, kept apart from this simulation. The method, the coverage checks and every simulated result are published with the backtest.