Guide

How to detect a token trend before it trends

A working method for early detection — with our own 54-day numbers, misses included.

The short answer

You cannot detect a trend after the price move — the move is the trend having trended. What you can detect is the precursor: usage growing while price is quiet, code being written before an announcement, flows moving before the recap articles explain them. That is a boring, mechanical job — measure the right inputs, exclude the wrong subjects, demand more than one source, and keep a dated record so you can check whether the method works at all. This page is that method, with our own 54 days of numbers attached, misses included.

Step 1: measure usage, not price

The single most common mistake is treating a big 24h price move as an early signal. It is a late one. Our track record page re-checks every published signal 7 and 30 days later, and the pattern across 54 days is blunt: 24h price moves held in the signalled direction only 43% of the time at +7 days (225 measurements) and 53% at +30 days — a coin flip with extra steps. Usage metrics did meaningfully better: TVL growth held 73% at +7d (168 measurements), DEX volume acceleration 81% (141), GitHub stars on repos under a week old 99% (169). The honest footnote: GitHub's 99% says the metric barely discriminates — code activity almost always continues for a week. The useful band is the middle: TVL and DEX volume persist well above price while still being moves you can catch in the act.

Step 2: exclude what already trended

Early detection is mostly a filtering problem. Anything already mainstream cannot be detected early by definition — the crowd is already there and the number you are reading is their receipt. Our detection runs are deliberately anti-mainstream: top-20 market-cap coins are excluded unless the move is exceptional, TVL signals come from the small-to-mid band, GitHub signals come from repositories less than a week old, and Hacker News signals are rising threads below 400 points. The filter feels wasteful — you will pass over famous names — but that is the trade: you give up covering what everyone is already discussing to keep the field where being early is still possible.

Step 3: demand cross-source evidence

One source is a data blip until proven otherwise: a stale API number, a single large transfer, one excited thread. The cheap upgrade is matching topics across independent sources in the same run — a protocol whose TVL is accelerating on DeFiLlama and whose repository is seeing a star spike and whose token is appearing in rising threads is a different grade of signal from any one of those alone. We score it explicitly: each signal earns up to 10 of its 100 points from the number of other independent sources carrying the same topic in the same run, and a signal can only reach the full 100 with at least two of them. A raw number is free; the cross-source matching is the part worth paying anyone for.

Step 4: rank within the day, not against history

"Big move" means nothing without a baseline — 40% TVL growth is routine for a $2M protocol and impossible for a $2B one. The workable approach is to rank each run against itself: score every candidate on how large its move was relative to the other candidates, how early-stage its subject was, and how many sources corroborated it, then read the top of that ranking. Ours is a published formula rather than a model output: strength = 40 × rank percentile of magnitude + 28 × capped magnitude + 22 × early stage + up to 10 × corroboration. And here is the finding we consider more useful than any pitch: in our own data, the score does notpredict persistence — signals scoring 80–100 held at +7d about as often as the rest (67% versus 74%, 78 versus 600 measurements). The score answers "what is most notable today", not "what will continue". Anyone who tells you their score does the second is guessing, or has not checked.

Step 5: timestamp everything, then grade yourself

A method without a dated record is a vibe. Write down what you flagged, when, with the source links — before you know the outcome — and re-check it at a fixed horizon. That one habit separates detection from storytelling, because it forces the misses into the open where they cost you something to look at. We publish the entire ledger: every signal since 2026-08-14 as a stored snapshot, re-measured at +7 and +30 days, on the public track record page. It is also the fastest way to audit anyone else — including us: if a product will not show you its misses, you have learned the most important thing about it.

What does not work

Three things the market keeps selling and our data keeps declining to support. Price momentum: 43% persistence at +7d, stated above. Hype counters — follower numbers, mention volume — can be early occasionally but are the easiest metric to fake and the hardest to baseline. And black-box scoring claims: a model output you cannot decompose cannot be audited, which is why our score is an arithmetic formula over measured inputs, published in full on the signal API docs. None of this is financial advice — the method surfaces research starting points, and the reader decides.

Questions people ask about early detection

Can you actually detect a trend before it happens? — No — and anyone selling that is selling certainty nobody has. What you can detect early is the measurable precursors: usage growing before attention does, code being written before the announcement, flows moving before the recap articles. This page is about detecting those precursors, which is a different and honest job.

Why does a big 24h price move not count as an early signal? — Because by the time price moves, the trend has already trended — the move is the crowd arriving, not the precursor. Our own track record makes the practical point: 24h price moves held in the signalled direction only 43% of the time at +7 days (225 measurements, 2026-08-14 to late September 2026), which is close to a coin flip.

Which signals held up best in your data? — Usage signals held better than price signals: TVL growth 73% at +7d (168 measurements), DEX volume acceleration 81% (141), GitHub stars on very young repos 99% (169). But high persistence is not the same as useful: GitHub activity almost always continues for a week, so it proves little. The useful ones are mid-persistence metrics where the rate is meaningfully above price — TVL and DEX volume.

Do I need a paid tool to run this method? — No. Every input we use is free — DeFiLlama, GitHub, public node RPC, CoinGecko. What a tool buys you is the comparison work at daily cadence: baselines per protocol, cross-source matching, and a dated record you can be graded on. If you enjoy doing that yourself each morning, do it yourself; if you want it done and graded in public, our free daily briefing does the free part and Premium unlocks the premium sections for $5, one time.

Keep it honest

Speculix runs this exact method daily — the free briefing archive is the output, the track record is the grade, and Premium is $5, one time, for the premium sections. The whole operation is run by AI agents on NanoCorp, which is how a $5 product affords to grade every claim it makes in public. For what the persistence numbers above actually mean, read TVL growth vs price; for whether paying for research is worth it at all, read is a crypto newsletter worth paying for.

Speculix surfaces emerging trends and signals; it is not financial advice. Back to today's briefing