Every change that rewired gas
"Gas price" sounds like one continuous series, but the machine generating it was redesigned repeatedly. The big ones for the math:
Genesis: gas as convention
Blocks are nearly empty; most wallets pay a hard-coded default (~50 gwei). Price reflects software defaults, not demand. A fee "market" barely exists.
Gas repricing after DoS attacks
Opcode costs rebalanced. Changes what gas buys, not how it's priced — minor for the indicator.
The fee market wakes up
Demand exceeds blockspace for the first time (ICO gas wars, then CryptoKitties). From here, gas price is a true first-price auction — users bid, the highest bids get in. Gas finally measures crowd urgency.
Supply inches up: 6.7M → 15M
Miners vote the block gas limit upward in steps. Each raise slightly deflates gas price for the same demand.
The auction is abolished
Fee formation is replaced by an algorithmic base fee that adjusts ±12.5% per block targeting half-full blocks — and is burned. Limit doubles to 30M. Same demand now produces a differently-shaped price series: smoother baseline, sharper spikes.
Proof of stake
Block times become a fixed 12s (+~3% throughput, more regular inclusion). Fee mechanics untouched — the smallest break of the four.
Blobs: activity stops paying L1 gas
Rollups move their data into blobs with a separate fee market. The largest source of L1 gas demand walks away overnight; user activity migrates to L2s where it no longer registers in L1 gas price at all. The biggest break in the series.
Pectra + relentless limit raises
Gas limit: 30→36M (Feb) → 45M (Jul) → 60M (Nov). Pectra (May 7) doubles blob capacity 3/6→6/9. Supply grows ~2× in one year while demand sits on L2s — structural fee deflation.
PeerDAS: the deflation continues
Blob capacity steps to 10/15, then 14/21; 80–100M gas limits proposed for 2026. Any indicator using absolute fee levels is permanently broken; even relative ones must adapt fast.
Four epochs, four different machines
Grouping by how the gas price was formed: E1 auction · immature E2 auction · mature E3 EIP-1559 burn E4 blob / L2. Two raw series below — the familiar per-transfer fee, and total network fee revenue (gas used × gas price × ETH price), which is more robust when the gas limit doubles.
Fee per transfer & total fee revenue, by epoch
The structural break, in numbers
| Epoch | Period | Mean fee | Structural drift | Time in high tide | Time in low tide |
|---|
Verdict: one formula does NOT fit all four epochs
Fees drifted +372%/yr in E1, +84%/yr in E2, then −52%/yr in E3 and −83%/yr in E4. The v1 z-score assumes a stable baseline, so in the blob era it sits in "low tide" 33% of the time and reaches "high tide" barely 1% — the indicator is being dragged by protocol engineering, not market mood. The 365-day rolling window self-heals within ~a year of each break (that's why v1 still worked), but every break costs it accuracy right when it matters. The math must adapt.
Ten candidate formulas enter
Each variant is scored against the 13 ground-truth tops & bottoms (hits within ±120 days, false signals) and on strategy equity with next-day execution. New in round 2: every variant is run across all 16 threshold combinations, and the table shows the median and worst equity of the whole grid — so a candidate can't win by cherry-picking its best cell. This table is computed live from the data, not typed in.
Results — all candidates, unedited, sorted by grid median
| Candidate | Idea | Events hit | False sig. | Best | Grid median | Grid worst |
|---|
The round-2 lesson: ridges vs plateaus
Round 1 crowned V3 (Revenue Tide) on its best cell. With three more weeks of data, V3's best cell still looks fine — but its grid median is below buy & hold. Its performance is a narrow ridge: step off the exact thresholds and the edge vanishes. V5 (Activity Tide) is a plateau: its worst threshold combination still beats holding, and it has the fewest false signals of any candidate. Same lesson as ever in backtesting: promote by the floor, not the peak. The old finding stands, refined: revenue level is a better fuel than per-transfer fee, but the real upgrade is the momentum leg — transaction count, which no fee-market redesign has ever rewired.
Is the tide secretly just reading ETH's price?
A fair objection: fees-in-USD contain the ETH price, so maybe the indicator is price momentum wearing a costume. Round 2 tested it. V6m strips every dollar out — revenue denominated in ETH, momentum from transaction count — and its grid median is ×1,073 vs — hold: the thesis survives with zero price input. But V7 (the raw gwei number alone) collapses to ×5 — gas price by itself is noise across regime changes. What carries the signal is economic weight (what users collectively pay) and participation (how many show up) — not the gwei readout.
Activity Tide (V5) — the new lead ★
Two ingredients, both on-chain: the level of total daily fee revenue (what the crowd is collectively paying) and the momentum of transaction count (whether the crowd is growing or thinning). Fee-market redesigns keep rewiring the price of gas — but none of them changed what a transaction is. That's why this one's backtest barely moves when the data updates, while price-based variants wobble.
ETH price · Activity Tide calls
Activity Tide vs the old champion vs v1 vs holding
Every V5 trade
| # | Bought | at | Sold | at | Return |
|---|
Spring 2026: the whipsaw nobody dodged
This lab opened on June 11, 2026 — right in the middle of the first real out-of-sample event: a bear-market rally that rolled over into a crash to $1,565. Here is what each system actually did, signal dates and prices, no edits.
Every system's 2026, side by side
| System | Bought | Sold | Round trip | The June crash to $1,565 |
|---|
Verdict: everyone lost — the difference is how much, and what came next
All three systems bought the spring rally and were wrong. But v1 and V3 bought late (late April) and sold into the June crash — v1's round trip was −30%, the worst trade in its recorded history. V5 was out by April 12 and sat in cash through the entire June leg down. A cycle indicator can't dodge every whipsaw; what it must do is fail smaller and stay out of the crash that follows. Round 2's promotion of V5 rests on three legs: highest floor across all thresholds, fewest false signals, and the best behaviour in the one live test we have. It still must call the next major turn live before it touches the homepage.
The Turn Confirmer (V4)
Detrended revenue z-score + momentum. It is not a strategy — it's a dated stamp that a major turn has happened. It confirmed 10 of 13 major tops & bottoms, typically ~6–7 weeks after the extreme — and nailed the COVID bottom same-day. Use case: "is this dip a cycle turn, or noise?"
Turn confirmations on the price history
Scorecard against all 13 ground-truth turns
| Actual extreme | ETH | Confirmed on | Delay |
|---|
⚖️ Lab rules — why this page isn't the homepage (yet)
- 13 events, 10 variants, grid-searched thresholds. That's a lot of searching over little ground truth. Round 1's champion proved the point: V3's best cell looked great, its grid median sits below buy & hold. The arena now reports median and worst across the whole grid, and promotion goes by the floor, not the peak.
- The 13th event is provisional. The June 25, 2026 bottom at $1,565 sits nine days from the edge of the data — if price breaks below it, it un-happens, and every "hit" on it un-happens too.
- The epoch problem is real but not fully solved. Tx-count momentum sidesteps fee-market redesigns, but Dencun moved activity to L2s where it doesn't register in L1 tx count either — and Etherscan publishes no blob-fee history to patch that with. An epoch-proof single formula remains open.
- E4 keeps mutating. Fusaka BPOs and 80–100M gas proposals mean the deflation isn't done. Every supply jump is a small fresh break.
- Same caveats as v1: daily averages, no fees/slippage in equity curves, not financial advice.