From the shadow on the stellar disc to the deviation from the ephemeris: the mathematics behind every analysis step, in seven chapters to play with. Each formula sits next to its animation and recalculates as you move the sliders.
As you scroll, each story step sets the animation to a matching example. Afterwards you can keep adjusting freely; clicking a step restores it. Where real measurements appear, they are marked as such; all other curves are model calculations.
Transit geometry
Four quantities set the shape of a transit:
They come out of the light curve through depth, duration and the length of the slopes.
At b = 0 the planet passes right through the centre of the star. The chord across the disc is longest, ingress and egress are short, and in between there is a flat bottom. Without limb darkening its depth is simply the area ratio.
The numbers 1 to 4 in the curve are the contacts: first and last touch of the discs from outside (1, 4) and inside (2, 3).
The further the path lies from the centre, the shorter the chord and the longer ingress and egress last. The duration from first to fourth contact follows from Pythagoras on the sky and the orbital angle:
In the concept's worked example (P = 3.0 d, a/R★ = 8, k = 0.10) the transit loses almost an hour from b = 0 to b = 0.8: 3.16 h versus 2.18 h.
At b = 0.95 the planet is never fully on the disc. There is no second and third contact, the curve becomes pointed and shallower than k². A V shape in measured data therefore does not automatically mean “small planet”.
The dashed curve belongs to b = 0.7 and a/R★ = 6.0, with its k fitted by least squares. It differs only in ingress and egress, and there by less than 2 ppt. With a few ppt of noise per data point the two are hard to tell apart.
That is why the pipeline uses a fixed geometry by default: a/R★ and i come from the literature, and only Tc, k and the systematics are free. Tc hardly suffers, because it depends mainly on the symmetry of the slopes.
Under “Example” you load rounded literature values of the measured targets. WASP-52b stands out with k ≈ 0.165: 27 ppt depth, almost three times as deep as HAT-P-33b. KELT-7b, with a 2.7 d period and a/R★ ≈ 5.5, is a close-in planet around a large star, so its transit is long.
From a/R★ and P, Kepler's third law even gives the mean stellar density, without distance or stellar radius:
Values rounded and for illustration only. An error of 5 % in a/R★ becomes about 15 % in ρ★.
Light curve model and limb darkening
At the edge of the star we look obliquely into the photosphere and see higher, cooler layers. The limb is darker, more so in blue than in red. A planet in front of the centre therefore blocks brighter light than at the limb.
Quadratic limb darkening describes the brightness as a function of μ = cos θ, which is 1 at the centre and 0 at the limb:
Top right shows the profile over the radius. The dashed line is the mean brightness of the whole disc, 1 − u₁/3 − u₂/6.
The relative flux is one minus the covered fraction of the intensity integral over the disc. Mandel and Agol (2002) solved this analytically; here it is integrated numerically ring by ring.
At the limb the planet blocks darker light, at the centre brighter light. The dashed curve below is the same planet without limb darkening: a flat bottom at k².
At mid-transit with b = 0 the following holds approximately
With u₁ = 0.40, u₂ = 0.25 the factor is 1.21. So k = 0.1058 gives about 13.6 ppt instead of 11.2 ppt.
Measurement: HAT-P-33b, night of 25/26 Feb 2026, R filter. 14.4 mmag were expected.
Switch between B and R. In blue the limb is darker, the bottom rounder and the centre deeper. If you report “Clear” instead of “R” to ExoClock, ExoClock assumes different u₁, u₂: the depth shifts by a few percent, while the mid-time Tc stays, because the curve stays symmetric.
The coefficients under “Filter” are rough guide values for a Sun-like star. The pipeline currently uses fixed u₁ = 0.40, u₂ = 0.25; values per star and filter after Claret are planned as an extension.
Aperture photometry
The aperture sums the starlight in a circle and subtracts the sky from an annulus. Its radius is a compromise: too small loses light when the seeing changes, too large collects sky noise.
On the left you see a cutout of the sensor around the target star, pixel by pixel. Atmospheric turbulence smears the star image to a typical full width at half maximum (FWHM) of 2–4″. Over the night it breathes: the bundle gets wider and narrower again.
An aperture never captures all the light, only the part of the point spread function inside the radius. For a Gaussian star image:
Up to about 1 · FWHM the fraction rises steeply, then it levels off. Below 1.5 · FWHM the flux reacts sensitively to every small centring error.
The CCD equation collects all noise sources in the aperture:
The dashed curve top right is the SNR versus radius. For faint stars the optimum lies around 0.7–1.0 · FWHM, for bright ones further out, because there the photon noise of the star dominates over the sky noise. Try the star signal slider.
With a fixed r, η drops when the seeing gets worse. In the ratio of target to comparison star this cancels as long as all stars have the same shape. If the star image varies slightly across the field (here: target 8 % wider), a residual remains that looks like a trend.
If you set r = k · FWHM for each frame, η stays constant and the residual vanishes. The depth is not biased, because target and comparison stars get the same radius in the same frame. The FWHM comes only from the comparison stars, smoothed, without the target star.
Measurements: For HAT-P-42b the noise with a fixed aperture (10 px) was 4.36 ppt, with a growing aperture (k = 2) 4.39 ppt: equivalent in calm seeing. For WASP-52b, in the run with dark, flat and bias, the growing aperture with k = 1.5 won.
Differential photometry and airmass
The atmosphere changes the brightness of all stars in the field almost equally. Dividing the target by a sum of comparison stars removes this. What remains is what only the target does, as well as the comparison stars resemble the target and are themselves quiet.
The night follows HAT-P-33b on 25/26 Feb 2026: the target crosses the meridian shortly after the start and sinks to airmass 2.3 by the end. The extinction in magnitudes grows linearly with the airmass X:
In the middle panel the target (coloured) and four comparison stars (grey) drop together. The transit is barely visible in the raw flux.
The ensemble is a sum of the comparison stars. The tool sums them unweighted, so brighter stars automatically count more:
At the bottom the transit remains, cleanly on a flat baseline.
Two hazy patches pass and swallow up to 25 % of the light. Because they hit all stars equally, they almost completely vanish in the ratio. Only slightly more noise remains, because fewer photons arrive in those frames.
Blue light is scattered more strongly. If the comparison stars are redder than the target, they become relatively brighter with increasing airmass, and a trend proportional to Δ(B−V) · X remains in the ratio. The pipeline models this residual as an airmass polynomial, together with the transit.
Pitfall from HAT-P-33b: if the airmass rises monotonically over the transit and the baseline before it is short (0.65 h there), the trend can swallow part of the depth.
A single star with share w of the ensemble and brightness change A biases the light curve by about w · A. If it gets brighter exactly in the transit window, the denominator grows and the transit appears deeper.
Measurement: For HAT-P-33b, G06 was variable according to Gaia and 7.2 ppt brighter in the transit window (7σ). With G06 the fit gave Rp/R★ = 0.0962, without G06 0.0927. The out-of-transit noise was the same in both cases (9.43 ppt), which is why the test “comparison star in transit window” has been a fixed part of the result sheet ever since.
The noise of the ratio contains that of the target and that of the ensemble:
With one equally bright comparison star the noise rises by 41 %, with five by 10 %, with ten by 5 %. More stars hardly help after that, but raise the risk of including a variable.
For HAT-P-42b three candidates were excluded: RUWE 23, Gaia flag VARIABLE, and one that changed by −4.2 ± 1.3 ppt in the transit window.
Noise
After the fit, residuals remain. Whether they scatter randomly or are correlated in time decides how honest the error bars are. For bright stars, moreover, the air often sets the noise, not the camera.
At the top you see residuals of a night, overlaid with means of n points each. If the noise is independent from point to point, the scatter of the means falls as 1/√n: in the lower plot the points lie on the line.
Seeing waves, tracking or thin clouds create fluctuations that are correlated in time. Here they are mixed in as a process with a correlation time of 25 min. The binned points now wander in waves, and the scatter falls more slowly than 1/√n.
Winn et al. (2008) compare the scatter of the bin means with the expectation for white noise:
β = 1 means white noise. The pipeline averages β over bin lengths of 10 to 30 min, the order of an ingress duration, and multiplies the error bars by it.
Measurement: HAT-P-33b gave β = 1.22 without flat and 1.04 with flat.
Atmospheric turbulence makes the brightness flicker, more strongly with a small aperture, short exposure and high airmass. After Young (1967):
With 150 mm aperture, X = 1.5 and 20 s this is about 4.5 ppt per frame, with the empirical Osborn correction rather 7 ppt. That is most of the measured 9.4 ppt from this night.
KELT-7 is so bright at G = 8.5 that exposures were 1.5 to 2.5 s. Scintillation then rises to about 25 ppt per frame, and the shutter was open only about 6 % of the time. The mid-time is correspondingly poorly determined.
Measurement: KELT-7b, 18 Oct 2026: O−C +25.5 ± 7.0 min after per-frame weighting. For the next bright star: expose 20 to 60 s and defocus slightly, roughly a factor of 3 to 4 less noise.
Fit and MCMC
For ExoClock not only the mid-transit time counts, but also its error: every O−C is measured against it. The MCMC (Markov chain Monte Carlo: a chain of random trial steps through the possible parameter values, in which each step depends only on the previous one) provides this error by trying out which mid-times fit the data. It has traps, though, and one of them is clouds at egress. The full theory is on the deep-dive page.
A fit finds the mid-time Tc and the radius ratio k for which the model best matches the points. That gives a value, but not yet an error. How precise Tc is depends on how wide the χ² valley around the best value is. In addition, parameters can compensate each other, and the valley can be skewed. A single error value from the fit often underestimates this.
How well a model fits is measured by χ² (“chi-square”): for each data point the distance between measurement and model is divided by the point's error, squared, and summed over all points:
Dividing by σi makes a point with a small error count more. The square makes positive and negative deviations equally bad and penalises large outliers especially. If the model fits within the errors, each point contributes about 1 on average, so χ² is roughly N. That is why the table on the right shows χ²/dof, which is close to 1 for a good fit (dof = number of points minus number of free parameters). Much larger means: the model does not fit or the errors were set too small.
From χ² follows directly the probability of each combination:
Top left shows χ² for all combinations. The shaded areas mark the 1σ, 2σ and 3σ regions.
The pipeline puts sixteen walkers into the valley. Each proposes a jump per step. If it leads to a smaller χ², it is always accepted; if it goes uphill, only sometimes, and the steeper, the more rarely. As a result the walkers stay at each place as often as it is probable. After a few thousand steps their positions form the distribution of Tc and k.
Beforehand, an optimiser searches for the best fit, starting from five Tc values within ±30 min. The walkers start there. The first steps count as burn-in and are discarded.
From the collected positions the pipeline reports the median and the 16th and 84th percentile:
Correlated noise contains less information than the number of points suggests. That is why the pipeline stretches all errors by β, and the distribution becomes wider by exactly this factor. Set here are β = 1.22 and 9.4 ppt per point as for HAT-P-33b. The simulation, however, has only 180 instead of 1077 points and therefore a larger error.
Measurement: HAT-P-33b, 1077 frames of 20 s, 9.4 ppt per frame, β = 1.22: Tc ± 2.2 min, R-hat 1.003.
Here the observation only starts at mid-transit, and shortly after egress a thin cloud pushes the flux down by 10 ppt for 10 minutes. A model with a later egress then explains the cloud as the rest of the transit. A second valley appears in χ² at ΔTc ≈ +25 min.
If the walkers land there, everything looks flawless: narrow histogram, normal acceptance rate, R-hat ≈ 1. Still, Tc is off by 25 min. No walker crosses the ridge between the valleys, and R-hat only compares walkers that all stand in the same valley. The MCMC measures the width of the valley it is in, but does not check whether it is the right one.
Measurement: For KPS-1b the points scattered at egress, and the ExoClock reviewer suspects exactly such a local minimum in the fit.
The pipeline's five start values already help: here the optimisers starting at −30, −15 and 0 min find the right valley, and the smaller χ² wins. But this only protects as long as the right valley is also the deepest. In the simulation the wrong valley was deeper for 5 out of 14 noise patterns. Try it with “New data”.
Check the profile: compute the best χ² for every mid-time (bottom right). A second valley with Δχ²/β² below about 10 is a warning sign.
Leave out the disturbed minutes: repeat the fit without the section at egress. If Tc jumps by more than its error, the result is not reliable.
Cross-check the depth: a side valley often pulls Rp/R★ away too, here to 0.090 instead of 0.100. The drift against the literature value from chapter 7 is therefore also a test for Tc.
How exactly the walkers jump, what R-hat and the autocorrelation measure, and how this looks on toy distributions is shown on the MCMC deep-dive page.
Ephemerides, O−C and TTVs
An ephemeris predicts when the next transit will happen. Every new measurement tests and improves it. That is exactly why ExoClock collects amateur data: from about 2029 the Ariel mission needs transit times predicted to within a few minutes.
Above are ten valid measurements, each against the ExoClock ephemeris of its planet:
Six nights lie within their 1σ error, HAT-P-12b and KPS-1b within 2σ. WASP-135b and KELT-7b are off by a good 3σ of their own measurement error. For WASP-135b the offset stayed stable between −2.5 and −4.6 min in all tests; whether the prediction itself is slightly off only shows once its own error is included. For KELT-7b the clock is ruled out (NTP log checked); a systematic in the noisy ingress is more likely.
| Planet | Night | O−C [min] | Rp/R★ | Drift |
|---|---|---|---|---|
| HAT-P-42b | 2026-01-17 | +3.7 ± 4.2 | 0.0812 ± 0.0079 | −0.56 |
| TOI-628b | 2026-01-19 | +1.6 ± 8.6 | 0.0724 ± 0.0103 | −0.84 |
| HAT-P-33b | 2026-02-25 | −0.1 ± 2.2 | 0.0932 ± 0.0038 | −3.16 |
| K2-115b | 2026-03-10 | −2.6 ± 3.9 | 0.1324 ± 0.0080 | +0.87 |
| HAT-P-18b | 2026-05-09 | +0.0 ± 1.2 | 0.1423 ± 0.0023 | +1.84 |
| HAT-P-12b | 2026-05-18 | +2.0 ± 1.1 | 0.1425 ± 0.0030 | +0.57 |
| KPS-1b | 2026-06-05 | −2.5 ± 1.8 | 0.1237 ± 0.0056 | +1.40 |
| WASP-135b | 2026-06-07 | −3.7 ± 1.1 | 0.1428 ± 0.0053 | +0.63 |
| WASP-52b | 2026-08-03 | −0.5 ± 0.6 | 0.1656 ± 0.0022 | +0.41 |
| KELT-7b | 2026-10-18 | +25.4 ± 7.0 | 0.0984 ± 0.0110 | +0.67 |
z is computed here only with the measurement error; the error of the prediction is missing in this plot.
The linear ephemeris and its error (T₀ and P treated as independent for simplicity):
With σP = 1 s, 17 min of uncertainty are reached after 1000 epochs, about 10 years at P = 3.47 d. From then on every measurement with ±2 min adds real information.
If the period is set 0.8 s too short, O−C grows linearly with the epoch. The most common cause of large O−C is exactly this: an outdated ephemeris whose error has added up over the years. A weighted straight-line fit through many measurements gives the corrected period.
Another planet, tidal decay or apsidal precession can shift the transit times periodically (transit timing variations). For most hot Jupiters they are smaller than a minute. Here a variation of 3 min is set: only many measurements with small errors make it visible.
ExoClock fits every curve itself and compares the depth with the literature. If it differs strongly, the curve contains systematics, and these could also have shifted Tc.
Below 2 is good, below 3 acceptable, from 3 on red. The slider below uses the errors from HAT-P-33b.
Measurement: HAT-P-33b: (0.0932 − 0.1058) / √(0.0038² + 0.0011²) = −3.16. The mid-time was solid, the depth did not pass the criterion.