Solar Wind Data Predicts Geomagnetic Storms
Four to Eight Hours Ahead

A satellite parked upwind

About a million miles sunward of Earth there is a point where the Sun's gravity and Earth's gravity balance, called L1. A spacecraft parked there stays put while the solar wind, the steady million-mile-per-hour stream of charged particles flowing off the Sun, washes over it on the way to us. The DSCOVR satellite (the Deep Space Climate Observatory, operated by the National Oceanic and Atmospheric Administration, NOAA) sits at L1 and reports the wind's properties every minute. It is, literally, a weather buoy moored upwind of the whole planet.

One of its readings matters more than all the others for storms: the north-south direction of the magnetic field carried inside the solar wind, a number called Bz. When Bz points north, the wind mostly slides around Earth's magnetic bubble. When it tips south, the two fields connect, and the wind's energy pours in. South is the key in the lock.

On the ground, the storm shows up as a ring of electric current circling the planet. Its strength is tracked by the Dst index, computed each hour from four magnetometer stations near the equator. A quiet Dst sits near zero; a storm drives it sharply negative. Dst is the number behind every "geomagnetic storm" headline.

The question

Every space-physics textbook says southward Bz drives storms. We wanted to test something more specific, in our own ingested feeds rather than anyone's summary: does the upstream needle actually predict the ground index, hour by hour, and with how much lead time? This is not a model or a forecast: it is two independent streams of measurements, one a million miles upwind of the other, laid side by side.

How we timed it

We took the 186 days where our DSCOVR and Dst records fully overlap (October 10, 2025 to June 30, 2026), which gives 4,469 hourly samples. Then we asked three questions. First, does knowing the recent history of Bz improve a prediction of Dst beyond what Dst's own history already tells you? That is a standard test called Granger causality, and we ran it at lead times of one, two, four, and six hours. Second, at what delay do the two series line up best? Third, and most important, does the whole pipeline stay quiet when we feed it deliberately scrambled data?

The link is unmistakable

At every lead time we tested, past Bz sharply improves the prediction of Dst. The test statistics range from 131 to 739, with the probability of arising by chance between 10⁻¹⁴⁴ and 10⁻¹⁶⁷. Those survive the standard correction for testing multiple lags by roughly forty orders of magnitude. And the physics-flavored version of the driver, Bz multiplied by the wind's speed (a stand-in for the electric field the wind drags past Earth), predicts slightly better than Bz alone, which is exactly what the textbook coupling theory says should happen.

The two clocks

The interesting part is not that the link exists. It is that the response runs on two clocks at once.

Think of the ring current as a bathtub. Southward Bz opens the tap. The rate at which water pours in responds fast: the hour-to-hour change in Dst tracks Bz with a lag of about one hour. But the water level, the storm's actual depth, keeps rising as long as the tap stays open. The level correlates best with Bz readings from four to eight hours earlier, peaking near five hours with a correlation of r = 0.42. That is not the wind taking six hours to arrive; the wind covers the last million miles in well under an hour. It is the storm integrating its driver, the bathtub filling.

One honesty note belongs here. Our written-down hypothesis predicted a sharp two-hour peak in Dst itself. That specific claim is not supported: Dst's own peak is the broad four-to-eight-hour plateau. The sharp roughly-one-hour response lives in the rate of change, the injection, and that is in fact the lag the classic storm equations (Burton 1975; O'Brien and McPherron 2000) predict. The physics passed the test; our phrasing of it did not.

Figure: the lag, drawn

Harder driving doesn't shorten the response

The storm equations also predict something more specific: stronger driving should shorten the response time. This is the one prediction the fuller record declines to confirm. Splitting the 4,469 hours by how hard Bz was pushing south, the coupling clearly gets stronger with harder driving — the peak correlation climbs from essentially zero in quiet hours (r = 0.02) to r = 0.21 under moderate driving to r = 0.47 in the strongly-driven hours (Bz below −10 nanotesla, now 144 hours, properly powered). But the response time does not shorten. The strongly-driven hours peak at about a five-hour lag — the same as the overall plateau — while the quiet hours peak nearer two hours but with almost no correlation to time at all. If anything, the harder the driving, the longer the storm integrates.

There is a cautionary tale in how this split first read. On the original 71-day window the strong group held only 45 hours and returned a nonsense eleven-hour lag — pure small-sample noise. With the record now more than doubled, that same group has 144 hours and lands cleanly at five hours. That both retires the spurious answer and removes the support for a "faster with stronger driving" reading: the timing-versus-strength prediction simply isn't borne out in this window. The main result — that Bz predicts Dst hours ahead, surviving every null — is untouched; it is only this secondary claim the data sets aside.

The control that keeps us honest

A pipeline that finds patterns should find nothing when there is nothing to find. We re-ran the full Granger analysis with the Bz series shuffled in time, destroying the real sequence while keeping every value. The shuffled driver failed at every lead time: zero of the four came out significant (all probabilities above 0.45). The machinery does not hallucinate causality.

What this doesn't show

• It covers 186 days, most of them quiet. The strongest events in the window were moderate storms. A longer record, which our ingestion builds automatically, will sharpen every number here.
• A simple straight-line fit of the storm equations to our hourly data explains about 10 percent of the variance, versus 40 to 60 percent in published storm-focused studies. That gap is mostly composition: published fits concentrate on storm intervals, while our window is dominated by quiet and recovery hours. A storm-only refit is queued for the follow-up.
• This is a statistical lead time, not an operational forecast. TerraPulse ingests measurements only; turning this lag into alerts is a different product with different obligations.

Why it matters

This lag chain is the physical basis of every aurora alert and satellite-drag warning in existence: a reading at L1, injection beginning within the hour, the storm peaking four to eight hours on. In earlier TerraPulse work we measured the next link downstream, an 80-minute median delay from the Bz excursion to the Kp index (the global geomagnetic disturbance scale) responding. Piece by piece, the Sun-to-ground timeline is showing up in our own feeds, with measurements at every step.

It is also the spine of what we are building next. TerraPulse's geomagnetic-storm monitor will watch this same Dst series live, detect storms against a frozen, pre-registered rule, and file a dossier of every sensor we run for each one. The lead time measured here is the reason a monitor like that gets to see storms coming at all.

Reproducibility

Every number on this page comes from the results.json in the bz-dst-predictive-lead-time workspace. Pipeline: 1-minute DSCOVR Bz and wind speed aggregated to hourly means and minima; hourly Dst from CDAWeb (NASA's Coordinated Data Analysis Web archive); stationarity checks on all series; cross-correlation at lags 0 to 12 hours with confidence bands from 1,000 phase-randomized surrogates; Granger causality at lags 1, 2, 4, and 6 hours with Bonferroni correction; a Burton-equation regression of the Dst rate of change on the drivers; a three-way split by driving strength with a boundary sensitivity sweep; and a strict shuffled-driver null.