From revenue recovery
to earnings per share.

Forecast issued 2026-11-01 · Revenue-to-EPS bridge · v3 · Data through 2026-06-30 · 26 revenue quarters · 22 post-IPO EPS quarters

Upstart reported $364.7M in Q2 revenue, up 41.7% year over year, and $0.16 diluted GAAP EPS. The lending recovery is visible; the durability of profitability remains harder to forecast. This revision connects the revenue forecast to contribution margins, fixed expenses, management’s profit outlook and dilution. The resulting GAAP EPS estimates can be compared with a dated, GAAP-labeled Street forecast. All three revisions precede the November 3 Q3 earnings release.

Our GAAP EPS scenario versus the Street

The comparison uses Finviz’s “GAAP EPS Performance and Forecast” table, checked 2026-11-01 with market data dated 2026-10-30. The quarter, currency and GAAP label match our target. Contributor dilution conventions are not individually disclosed, and some historical provider actuals differ from issuer diluted EPS. This is a GAAP-labeled comparison, not a fully audited reconciliation of every analyst estimate.

The last column substitutes the Street’s entire revenue path into our accounting bridge, holding the other assumptions fixed. It helps separate revenue differences from margin and expense assumptions; it is not the Street’s own earnings model. Street revenue estimates are $364.01M, $384.46M, $403.02M, $469.82M, in quarter order.

Reading the gap: Q3 is modestly below consensus. Q4 is above it because our revenue allocation is stronger and secured contribution margins reach breakeven. The lower 2027 EPS path reflects our lower revenue forecasts and continuing fixed-cost growth; management has not supplied the quarterly 2027 margin path used here.

Revenue and earnings forecasts

Revenue is in millions of dollars; EPS is diluted GAAP dollars per share. Q3 has already ended but its results are unreported as of this vintage. All observed quarters through Q2 enter the final fit. The guidance update applies to Q3 and Q4 2026 together. Revenue estimates and ranges retain revision v2. The new EPS path is an accounting transformation of those revenues and the assumptions below. Monthly originations remain outside this model.

Next four unreported quarters: $1.48B in modeled revenue, +18.7% versus the preceding four quarters. This total spans Q3 2026–Q2 2027; it is not a fiscal-year forecast. Quarterly ranges cannot be added to produce a calibrated annual range.

What management’s guidance adds

The August 4 outlook targets $1400M of full-year revenue. Subtracting the $672.922M already reported leaves $727.078M for Q3 and Q4 together. It supplies no separate quarterly allocation.

H2 aggregate uncertainty: approximate 80% range $425.7M–$1327.6M; 95% range $163.9M–$3728.0M. These ranges are calculated for the total, not by adding quarterly endpoints.

Management’s comparable guidance history

For each completed year, we checked the outlook issued with Q2 results. Only 2021 and 2025 supplied numerical consolidated full-year revenue guidance comparable to the current outlook. We subtract known H1 revenue and compare the implied H2 forecast with the first reported full-year result less H1. Quarterly guidance errors and fee-revenue guidance do not enter this calibration.

One outcome beat the implied H2 guidance and one missed. Mean absolute percentage error, measured against actual H2 revenue, was 10.2%. The geometric-average actual/guidance ratio implies a 9.7% upward adjustment. Two observations do not establish that this bias will persist.

How the two signals are fused

We use a regularized, accuracy-weighted linear opinion pool: a mixture of the model and calibrated-guidance forecast distributions. Management receives 60.6% of the mixture weight and the model 39.4%. This combines forecasts that may share information; it does not multiply them as independent evidence. A mixture weight is not a probability that management is correct.

The accuracy scores use historical mean squared log forecast errors, with a 1 + 1/n small-sample adjustment. The model is scored on 7 complete rolling two-quarter predictions; management is scored on 2 midyear outlooks. These samples differ in dates and market conditions. Both scores use the original, uncorrected forecasts so the fitted guidance bias cannot improve its own weighting score.

Raw inverse-error weighting would give guidance 81.7%. We shrink that weight toward 50% using λ = n / (n + 4), where n = min(n_model, n_guidance) = 2. The strength of four is a disclosed regularization choice, not an empirically optimal parameter. The comparison below shows how much the central result depends on weighting.

The H2 median is allocated 46.2% to Q3 using the original model’s Q3/Q4 ratio. The quarterly points add to the H2 median. They are central allocations, not necessarily the marginal quarterly medians. Every simulated Q3/Q4 pair adds to its own H2 total, while the historical variability of the ratio supplies timing uncertainty.

Preserving the original forecasts

All three November 1 vintages remain archived: v1 is the original quarterly model, v2 adds revenue-guidance fusion, and v3 adds the earnings bridge. Revenue remains at v2; v1/v2 EPS was $0.124 at every horizon. Later results will be scored against each saved version.

The business behind the numbers

Q2 fee revenue was $348.0M, or 95.4% of consolidated revenue. Net interest income and fair-value adjustments contributed another $16.7M. Revenue therefore reflects both lending activity and balance-sheet economics; loan volume alone does not determine earnings. Q2 earnings release.

The revised FY2026 central estimate is $1446.4M: Q1 and Q2 actuals plus the fused second-half estimate. Management’s August outlook was approximately $1,400M. The revised difference is 3.3%, with substantial uncertainty around it.

How the underlying quarterly model is tested

We compare seasonal growth, a fixed log trend with quarterly effects, and smooth time-varying trends against two simple baselines: repeat the latest revenue, or repeat the same quarter a year earlier. The underlying quarterly-model choice is log_trend. Its original point forecasts appear in the comparison above.

Dynamic candidates penalize second differences in the log trend with strengths 16, 64 or 256. Their forecast growth deviations damp toward the recent eight-quarter average by a factor of 0.8 per quarter. Seasonal growth uses the mean of the last eight year-over-year log growth rates. These are predictive time patterns; they do not identify a causal effect of AI, marketing or R&D.

At every forecast origin, an inner chronological comparison scores all available one-to-four-quarter outcomes using only earlier observations. Revenue selection minimizes mean absolute log error. It starts after 16 observations and uses at most 12 origins. The final selection repeats this process on the full 26-quarter history before refitting.

Reconcile the reported $0.16 EPS

Q2 operating income was $14.64M, a 4.0% margin. Diluted EPS uses a $17.070M numerator and 109.721M weighted-average diluted shares.

Tax is a deduction. The $0.531M convertible-note adjustment increases the diluted numerator above $16.539M of net income. Contributions sum to the pre-rounding EPS result, which rounds to $0.16. Q2 filing, earnings-per-share note.

How revenue becomes diluted GAAP EPS

The bridge starts with forecast consolidated revenue. The remaining annual fee-revenue guide implies a 92.83% fee share. We retain Q2’s unsecured/secured fee mix and unsecured contribution margin. Management expects secured contribution margin to reach breakeven in Q4 and fixed expenses to grow at a low single-digit sequential rate. We interpolate halfway to secured breakeven in Q3 and use 2% quarterly fixed-cost growth, with 1% and 3% sensitivities. Continuing these assumptions into 2027 is our scenario, not management guidance.

Contribution profit plus net interest and fair-value revenue, less fixed operating expenses, gives operating income. Those fixed expenses already include stock compensation and depreciation. Other income uses the recent four-quarter median ($1.735M), limiting the influence of one-time debt gains. Tax is the larger of a $0.325M recent-median expense floor and 1.81% of positive pretax income. Losses receive no assumed tax benefit. A 25% profit-tax sensitivity makes the low current tax rate explicit.

Basic shares grow at half the recent median quarterly log growth. The Q2 option/award increment of 4.783M shares and aggregate convertible increment of 8.364M are carried forward as assumptions. Losses use basic shares. For profits, the model adds the $0.531M conversion numerator adjustment and convertible shares only if doing so reduces EPS. Actual conversion terms, tranche ordering and stock-price-dependent option dilution may differ; the additional-share scenario is a sensitivity, not a share-price forecast.

A soft constraint from management’s profit outlook

The annual adjusted EBITDA guide is $294M. Deducting $117.374M already reported leaves $176.626M for H2. Dividing by the $727.078M implied H2 revenue guide gives a 24.29% H2 EBITDA margin. We use this margin as a soft constraint on the revenue-conditioned bridge. The annual 21% EBITDA margin is never substituted for a quarterly GAAP net margin.

To calibrate the constraint, we compare each eligible Q2 outlook’s implied remaining-year EBITDA margin with the realized H2 margin. Comparable numeric annual profit guidance exists for the same two years, 2021 and 2025. The older EBITDA definition differs, and conditioning on guidance being issued omits adverse periods without comparable outlooks.

We shrink the mean historical margin error toward zero by λ = n/(n+4) = 1/3, producing a 3.56 percentage-point adjustment. The calibrated target is 27.85%. Historical error scales and the same small-sample penalty give guidance only 8.32% of the reconciliation. The bridge determines the rest. This profit-constraint weight is distinct from the 60.6% revenue-mixture weight.

Assumptions that move EPS

Does the earnings bridge predict better?

Not established. The retrospective accounting core performs worse than both the old independent EPS model and repeating the latest reported EPS. The current management overlay has too few comparable historical vintages for an independent end-to-end test. This revision makes the economic assumptions auditable; it does not establish an earnings-surprise trading edge.

The accounting core carries the latest contribution margin and fee mix, projects fixed expenses and shares with half the recent median log growth, and includes other income, taxes and profit-sensitive dilution. A second version averages four quarters of margins/mix; the selection column chooses between them using only earlier outcomes. Inner tests start after eight EPS quarters; outer tests start after fourteen. Revenue is forecast afresh using only earlier observations, never replaced with realized revenue in this accuracy table. The current guided bridge instead uses the disclosed management assumptions. An earlier operating-income regression attempt also failed and remains in the research archive.

Quarterly-model validation and fusion limits

The following historical errors evaluate the original quarterly model. They do not establish the accuracy of the new fusion procedure. There are too few comparable guidance years for an independent end-to-end fusion backtest. The weights, guidance-bias estimate and intervals are estimates from small, different samples.

Outer tests begin with 18 revenue observations and 14 post-IPO EPS observations. Targets run from 2024-09-30 to 2026-06-30. At each origin the model is chosen afresh from earlier outcomes; the table evaluates that complete selection procedure, not just today's winner.

The unchanged 2027 revenue ranges use horizon-specific nested log errors. The prior v1/v2 EPS ranges used dollar errors; v3 replaces them with the accounting-scenario ranges described above. The root mean square error includes bias and is multiplied by a Student-t quantile with n−1 degrees of freedom and √(1+1/n). Revenue ranges are exponentiated; revenue point forecasts are on the log/median scale, not bias-adjusted means. Both 80% and 95% revenue ranges are preserved.

How this forecast will be scored

Keep all three November 1 revisions unchanged: v1 is the original quarterly model, v2 adds revenue-guidance fusion, and v3 adds the GAAP earnings bridge. At each release, record the first reported consolidated revenue and diluted GAAP EPS, their source and publication date, signed errors, revenue absolute percentage error, EPS absolute error, and coverage of the saved 80% and 95% ranges. Q4 EPS must come from a direct quarterly disclosure. Later restatements and new model vintages belong in separate records.