A Story That Fits Every Crash
On 12 March 2020, as covid lockdowns spread across Europe, the Nasdaq Composite fell 9.4%. The same day, bitcoin fell about 40%. The pattern repeated through 2022, when both assets bled red as the Federal Reserve tightened. Each time, commentators reached the same shorthand: bitcoin was tech beta with extra leverage, a risk-on amplifier of the Nasdaq's mood.
The story is satisfying because it explains every move. Stocks down, bitcoin down: risk-off. Stocks up, bitcoin up: risk-on. The framing is unfalsifiable in either direction. It is also wrong about what matters.
The Machine Disagrees
Revision note (June 2026). The first version of this article claimed the pipeline gave equities "zero feature importance," resting on a single early run. An audit flagged the claim as under-supported, and a dedicated re-run (v3: the current 25-feature set plus four equity features, full SHAP tables persisted, available as ml/results/walk_forward_v3.json) was carried out to settle it. The re-run contradicted the zero claim, and this section now reports what it actually found.
The v3 run trained gradient-boosted trees (XGBoost) on 4,323 daily observations and 29 features, including the S&P 500 and Nasdaq Composite in level and 7-day-return form, across four expanding walk-forward windows and three prediction horizons: twelve models. SHAP, a game-theoretic decomposition that attributes each prediction to individual feature contributions, then ranked the inputs. Put plainly, SHAP is a bookkeeping method that splits each prediction into credit and blame for every input, the way a shared bill is split by who ordered what.
Equity features are not zero. Across 48 equity feature-round observations (four equity features, each scored in twelve models: 48 chances to matter), they landed in a round's top five 4 times and in the top ten 21 times. The single best showing was the Nasdaq at rank 4 in the 7-day model of the 2022-2023 test window, the healthiest model in the set. In absolute terms the contributions are small, an order of magnitude below the round leaders in most cases, and comparable to what the dollar index and oil contribute: macro context, surfacing when the macro regime dominates.
The pattern in where they surface is the informative part. The equities' best ranks cluster in the 2022-2023 window, the one period in bitcoin's history when Federal Reserve tightening visibly drove both assets, and the period that produced the famous 0.7 correlation readings. In the 2017 retail mania and the 2019-2020 windows they sit mid-pack or lower, and the lagged-return variant (the same index expressed as its return over the prior seven days, a check on whether stocks lead) ranked dead last in 10 of 12 rounds. Equities show up exactly when the macro tide lifts or drains everything at once, and not otherwise. That is what a downstream co-mover looks like, not an upstream driver.
One caveat applies to v3 as much as it did to the earlier runs: most of the twelve models stopped at a handful of trees, so individual SHAP rankings are weak evidence (the SHAP analysis now documents this in detail). The equity result is best read at the level where it is consistent: never leading, strongest in the macro-driven window, redundant with the macro features already on the dashboard.
What the Bank of Spain Found
This is not a quirk of one model. In 2023 the Bank of Spain applied a comparable LSTM and SHAP framework to a different dataset and a longer horizon. Its working paper reported substantial contributions from technology variables, attention indicators, and on-chain metrics, with sharply different importance shares across cycles. Equity-market indices do not figure among the drivers it reports. That is corroboration of a softer kind than a head-to-head test (the two studies used different feature sets), but the direction agrees: when researchers decompose what predicts bitcoin, stock indices are not what surfaces.
Co-Movement Is Not a Signal
Pundits respond to this with rolling correlation charts. The 90-day correlation between bitcoin and the Nasdaq did breach 0.7 in mid-2022, the highest reading in the asset's history. Surely correlation that strong has to carry something predictive.
It does not. Correlation measures whether two series move in the same direction at the same time. It says nothing about which moves first. A flood that drowns both a house and the dog inside it produces a perfect correlation between the dog's altitude and the water level. Neither is predicting the other. Both are predicted by the rain.
For bitcoin and tech stocks, the rain is global liquidity. When the Federal Reserve eases, both assets receive flows. When it tightens, both lose them. The cross-asset correlation is real, but the driver is upstream. Once liquidity (captured on the dashboard as M2 supply, the broad tally of dollars sitting in accounts and cash, and the dollar index) sits in the model, the Nasdaq adds nothing on top.
A Softer Channel
None of this rules out an indirect effect. A booming equity market can draw attention and capital away from bitcoin without ever leading its price: when tech stocks deliver outsized returns, marginal investor flows and media coverage rotate towards them, and the asset that requires more explanation gets less of both. The 2026 backdrop is a live example. The S&P 500 and the Nasdaq have spent the year at or near record highs, and crypto allocations have visibly cooled even as bitcoin's own fundamentals held.
The effect is real but indirect. It works on the appetite of the marginal buyer, not on bitcoin's valuation, and most of it is already absorbed by the dollar index and M2 supply, which respond to the same liquidity backdrop that lifts equities. That absorption is the subject of its own piece: liquidity does not rank low because it does not matter, but because it is already counted. The model's silence on the Nasdaq is therefore not a claim that equities are irrelevant to investor behaviour. It is a claim that whatever they contribute, they contribute through channels already on the dashboard.
What Stays Off the Dashboard
The thirteen indicators in the composite were selected using SHAP rankings, economic rationale, and an independence filter (see the correlation analysis for the independence test). Equity indices fail the test that matters: redundancy. Their modest SHAP record is the profile of a macro echo, and the macro dimension is already covered by DXY and M2, with which any equity contribution overlaps. Adding the Nasdaq would re-buy information the dashboard already owns.
That is useful to internalise when reading the signal. When stocks crash and the dashboard barely moves, that is not a malfunction. It is the model behaving as designed, ignoring noise that looks important in real time and waiting for inputs with predictive history.
The reverse is also true. A roaring tech rally that fails to drag valuation indicators down does not weaken the case for accumulation. The Nasdaq's mood has no vote.
What It Means for Position Sizing
Treating bitcoin as tech beta carries practical costs. Investors size positions assuming the two assets share risk drivers, then add tech-stock hedges that do not hedge the risks they came to manage. A short-Nasdaq leg (a side bet that tech stocks will fall) can offset the second-order liquidity move that pushed both assets, but it leaves bitcoin's idiosyncratic risks (regulatory shocks, exchange failures, miner capitulations) untouched. The covariance (the co-movement) that does exist is macro liquidity, already on the dashboard: sizing against tech-beta is sizing against an echo the dashboard tracks directly.
The Inverted Question
The more useful question is not why equity indices fail to predict bitcoin. It is what does. SHAP's answer, drawn from three complete cycles and broadly consistent with the Bank of Spain's independent work, is that structural valuation (Power Law Position, 200-Week MA) and on-chain demand (BTC Fees) carry the most consistent signal. These features describe bitcoin's relationship to its own price history and its own usage. They do not depend on whether the S&P had a good morning.
That hierarchy was the first useful output of the prediction pipeline. The prediction model itself failed, averaging 49% directional accuracy (best LSTM configuration: 54%). The feature ranking was more useful. Among its findings, the marginality of equity indices was the most counter-intuitive and, after a re-run forced this article to soften "zero" to "mid-pack echo," the most carefully tested. See the full SHAP analysis for the broader feature ranking.
The Nasdaq has little to tell bitcoin that DXY and M2 do not already say. The dashboard is built accordingly.