stablecoin_share_anchor

1. It is a price chart, not a flow gauge — measured, at both horizons

The sales pitch is that a high share means capital has moved to the sidelines. Test it against the actual size of the stablecoin pile and it collapses. Across the 1,644 paired readings we hold, the correlation between the share and the total crypto market cap is −0.9964 (r² = 0.9929). The correlation between the share and the implied dollar value of the stablecoins themselves — the thing it claims to track — is −0.0662. Over that week the stablecoin pile moved 0.376% while the market cap moved 2.562%.

The obvious defence is that a week is far too short: issuance is slow, so of course the numerator looks frozen. So we tested it over the full reconstructed year, and the defence mostly fails. Day to day, the denominator is 11.3× as volatile as the numerator; the correlation between daily changes in the share and daily changes in the market cap (inverted) is 0.9961, r² 0.9923; the variance of daily supply changes is 0.81% of the variance of the share. Over the year as a whole the share nearly doubled, 5.699% → 11.131% — and with the supply frozen at day one it would still have travelled to 10.022%, which is 81.3% of the same journey. Stablecoins were real but minor: 18.7% of the move.

The concrete version: over that year the stablecoin supply grew 13.9%, while the crypto market fell 43.1%. Read as our copy tells you to read it, the "dry powder on the sidelines" roughly doubled. Almost all of that was the sidelines staying put while the field shrank.

2. We count two of the world's 416 stablecoins — and the denominator counts all of them

The poller keeps a list of 13 stablecoin tickers to look for. It finds them in CoinGecko's market_cap_percentage map — which only ever contains the ten largest coins in crypto overall. Two stablecoins are big enough to make that cut. The other eleven tickers on our list contribute exactly zero, every cycle, including Sky Dollar ($6.63B) and Dai ($4.79B), which are simply too small to be top-ten coins.

DefiLlama tracks 416 stablecoins worth $308.10B. We see USDT ($183.03B) and USDC ($72.20B) — $255.23B. We are blind to $52.86B, 17.2% of the stablecoin universe. And the two sides of the fraction do not match: the total market cap in the denominator is computed across all 18,377 coins CoinGecko lists, stablecoins very much included. So the ratio divides two-stablecoins-worth by an-everything-worth.

This is not cosmetic — it straddles our own recipe. Our reading was 11.2601%. The all-stablecoin share was 13.5947%, an understatement of 2.33 percentage points. The builder's one-click recipe is "share above 12%". On the real stablecoin universe that condition was true; on the number our gate actually reads it was false. The recipe we ship is switched off by our own measurement error. (That 13.59% figure divides DefiLlama's universe by CoinGecko's total, so it mixes two vendors and should be read as "about 2 pp", not to four decimals — but the direction and rough size are not in doubt.)

There is a second-order hazard in the same mechanism. Because membership of that map is decided by rank, a stablecoin crossing into or out of the top ten would move our number by billions overnight with nothing happening in the world. In fairness, it hasn't happened during our record, and it isn't close: USDC sits fifth, and the tenth-placed coin is roughly 2.7× larger than the biggest stablecoin outside the map.

3. The polarity is genuinely contested, and we assert one side as settled fact

The module, the builder and the starter bot all state the same reading without qualification: high share = dry powder = risk-off. There is a well-established opposite reading. New money entering crypto arrives as freshly minted stablecoins before it buys anything, so a rising absolute stablecoin supply is widely treated as incoming liquidity — constructive, not fearful. Both stories are coherent, they point opposite ways, and nothing on our site tells you the question is open.

There's a mechanical detail underneath that makes the "cash moving to the sidelines" story weaker than it sounds. Stablecoin supply changes only when coins are minted or redeemed. A trader who sells Bitcoin for USDT on an exchange moves existing tokens between wallets; total supply is unchanged. So the very act the pitch describes — investors retreating into cash — does not move the numerator at all. It moves the share only through the denominator, by pushing the price of what they sold down.

We are not going to tell you which reading is right, because we cannot measure it. We tried. Over the reconstructed year, a high share does associate with better forward returns (correlation +0.33 at 30 days, +0.70 at 90 days) — but running the identical test on 1 ÷ market cap, which contains no stablecoin data whatsoever, gives a stronger result (+0.38 and +0.75). The signal adds nothing over simply knowing the price had fallen. And at 90 days, 363 overlapping observations amount to about three independent windows, which is no evidence at all. The honest answer is that this question cannot be settled from what we keep.

4. One shipped setting is always on, another is almost never on — and our eight days can't tell you which

Measured against every reading in our table (range 11.0384% – 11.3273%):

On our own data every one of these is a constant — permanently open or permanently shut. A switch that can't change state isn't a switch.

But the year column is the fair answer, and it mostly exonerates the defaults. Over 363 days, "above 10" was open 44.9% of the time — close to an ideal duty cycle for a regime filter — and "below 10" the complementary 55.1%. Those are good, well-chosen thresholds. They look broken here only because our eight days happen to sit entirely between 11.04% and 11.33%, above one line and below the other. The defect is our memory, not the number.

The exception is the recipe. "Above 12" cleared on 1.7% of days in the last year — about six days out of 363, and given catch #2 it needs a real share nearer 14.3% to fire. That one is badly calibrated on any horizon, and it is the setting attached to the friendliest one-click button we offer.

5. Seven days of memory for a number with a multi-year arc

The poller prunes crypto_dominance to seven days. Over the last year this series ran from 5.424% to 12.179% — a range of 6.76 percentage points. We hold 0.29 of them, about 4% of the year's travel, and nothing older. Every question worth asking about a regime gate — how often does it open, how long does it stay open, is now unusual — is unanswerable from our own storage.

The prune also keeps slightly more than it advertises, the same way its two siblings do: it compares an ISO-8601 timestamp against SQLite's datetime() as a string, and because 'T' > ' ', rows from the cut-off day survive. Our window is 188.25 hours, not 168. Harmless here, and not fixed — this loop doesn't edit pollers.

6. The backtester can't replay it — and the code doesn't even know the signal exists

The string stablecoin appears zero times in quants/backtest.py. It is not in the supported list, and it is not in _UNSUPPORTED_ENTRY either — the documentary list of things we know we can't replay. Only the catch-all check saves it. A backtest containing this signal runs without the gate and still returns an equity curve; it flags unsupported_rules, but the curve you're looking at is a different strategy from the one you built.

The standing justification in the code is that these anchors "need a forward-captured series we don't have retroactively". Here that is half true, and worth splitting. The numerator is free and keyless back to 2017 — 3,178 days of it. The denominator genuinely is hard: CoinGecko's historical global market cap is paywalled, and we could not find a clean free keyless multi-year total-crypto-cap feed. So the excuse holds for the ratio as defined, and does not hold for the ingredient that actually carries the information.

7. Blind for more than a quarter of the clock, by design

Over our window the poller should have written 2,260 readings and wrote 1,644 — 72.7% coverage, 27.3% blind. The cause is the isolated try/except wrapped around the whole dominance step. That isolation is correct design — a CoinGecko hiccup must never break the core coin-price polling — but its consequence is that this row is the designated sacrifice whenever the database is contended. The poller's lifetime log carries 4,602 database is locked errors, 926 dominance-step failures, 434 fetch failures, and vendor losses of 112× HTTP 429 plus assorted 400s and gateway errors. (Those are lifetime counts; the log carries no timestamps, so they can't be attributed to a window.)

In fairness, missing a five-minute sample costs almost nothing for a variable this slow, and gaps never approached the failure threshold — see the defence below.

8. Configurations that validate but cannot mean anything

validate_config checks each bound is a number from 0 to 100 and never checks them against each other. So {"above": 12, "below": 8} passes validation and can never be true — the same defect documented on finance_anchor and both siblings on this API call. {"above": 100} and {"below": 0} are also legal and permanently shut.

Worse, to_english renders that impossible pair as two confident opposite statements joined by "and": "only when stablecoin share > 12% (lots of dry powder on the sidelines — risk-off) and stablecoin share < 8% (capital fully deployed into crypto — risk-on)". The genuinely useful configuration — a band, like {"above": 10, "below": 12} — works correctly but reads back as risk-off and risk-on simultaneously. Both comparisons are strict, so a threshold set exactly on the reading is a miss.