Brand Forecast Studio · a commercial data science case study by Mohammad Uzair Khan, on public US Medicaid data
Quarterly forecasts for 19 MSD brands across 6 business units: $1.9B of 2026 US Medicaid spend, against $1.9B in 2025
One machine-learning model, trained across 535 brands, forecasts each brand’s units four quarters ahead; price per unit follows the brand’s own recent trend. Tested on rolling origins from 2023Q4 to 2025Q3, it misses MSD brands’ spend by 17.9% on average, against 19.4% for the best simple baseline, and wins on 7 of 19 brands. Its P10 to P90 range is calibrated, and held the actual 80% of the time on quarters the calibration had not seen. Change price, volume or a generic launch below and the forecast moves with it, or ask a question in plain words.
Forecast and what-if
Pick a business unit and a brand. The yellow cells are assumptions: change them and the scenario line and table follow. Nothing here is sent anywhere; the arithmetic runs in this page.
JANUVIA: $227M forecast for 2026
Diabetes · US Medicaid spend by quarter, before rebates. The band runs from the P10 to the P90 forecast.
This is the model’s forecast. Change a yellow cell to build a scenario.
actual forecast P10 to P90
By business unit: 2 of 6 grow in 2026
Each unit’s total is the sum of its brands’ forecasts. Ranges are left out here: single brands’ P10s do not add up to a P10 for the unit. A scenario built above shows as a yellow mark on its unit.
US Medicaid spend by business unit, 2025 actual and 2026 forecast
Before rebates. Outlined bar: 2025; solid bar: the 2026 forecast (P50).
2025 actual 2026 forecast
Ask the forecast
Each question becomes one call to a typed tool (forecast, scenario, business-unit total, accuracy, ranking, health, range), and the answer comes only from the numbers on this page. The tool call is shown under the answer. A question it cannot place is declined rather than guessed. It runs in this page, with no language model and nothing sent anywhere.
Tested before it is trusted: 17.9% average miss on MSD brands, against 19.4% for the best baseline
Each origin from 2023Q4 to 2025Q3 trains only on what was known then and forecasts one to four quarters ahead. The miss is weighted absolute percentage error (WAPE): total absolute error over total actual spend. Across all 535 brands the model misses by 11.8%, against 12.8%. It is not best everywhere, and the table says where. The last two columns count forecasts made from 2024Q4 on, which the range’s calibration had not seen; four quarters ahead has none yet for MSD brands.
| Quarters ahead | Model | Last quarter | Same quarter last year | Recent trend | Inside the range, raw | Inside, calibrated |
|---|---|---|---|---|---|---|
| All | 17.9% | 21.5% | 26.9% | 19.4% | 70% | 80% |
| 1 | 9.8% | 10.9% | 20.1% | 10.4% | 74% | 80% |
| 2 | 15.9% | 17.7% | 24.1% | 16.5% | 61% | 79% |
| 3 | 22.5% | 27.2% | 30.7% | 24.7% | 79% | 79% |
| 4 | 28.9% | 37.7% | 37.7% | 32.5% | n/a | n/a |
| Brand | Business unit | Model | Best baseline | Best method |
|---|---|---|---|---|
| VERQUVO | Cardiovascular | 5.9% | 8.4% | Model |
| WINREVAIR | Cardiovascular | 15.9% | 47.0% | Model |
| JANUMET | Diabetes | 25.6% | 28.0% | Model |
| JANUVIA | Diabetes | 27.0% | 26.0% | Recent trend |
| SEGLUROMET | Diabetes | 9.6% | 5.0% | Recent trend |
| STEGLATRO | Diabetes | 7.2% | 9.4% | Model |
| DELSTRIGO | HIV | 6.8% | 4.9% | Last quarter |
| ISENTRESS | HIV | 4.7% | 6.9% | Model |
| PIFELTRO | HIV | 4.3% | 2.4% | Same quarter last year |
| BRIDION | Hospital & specialty | 43.4% | 35.9% | Same quarter last year |
| DIFICID | Hospital & specialty | 26.1% | 25.4% | Last quarter |
| EMEND | Hospital & specialty | 46.2% | 42.8% | Recent trend |
| PREVYMIS | Hospital & specialty | 9.3% | 9.5% | Model |
| RECARBRIO | Hospital & specialty | 63.2% | 52.3% | Last quarter |
| SIVEXTRO | Hospital & specialty | 25.0% | 24.0% | Last quarter |
| ZERBAXA | Hospital & specialty | 7.2% | 6.9% | Recent trend |
| BELSOMRA | Neuroscience | 5.9% | 4.6% | Same quarter last year |
| KEYTRUDA | Oncology | 14.9% | 12.4% | Recent trend |
| WELIREG | Oncology | 12.7% | 15.7% | Model |
Pipeline health: 9 checks pass, 4 need review, 0 fail
Run after every refresh, before a forecast is shared. A review is not a failure: it names what a person should look at, such as a price cut or a brand the model does not beat.
Method and limits
- Data. 13,336 brand-quarters for 535 brands with at least $20M of Medicaid spend in 2025, and every MSD brand above $2M, from the CMS State Drug Utilization Data cleaned in Rx Market Pulse.
- Model. One LightGBM model across all brands forecasts the change in log units from lags, recent growth, brand age, quarter, therapy area and price trend; two quantile models give P10 and P90. Price per unit follows each brand’s own trend over the last year, capped at about 5% a quarter. Spend is units times price.
- Range. The quantile models’ P10 to P90 band held the actual only 70% of the time for MSD brands. It is calibrated by conformal prediction: each band is widened by how far earlier actuals at the same horizon fell outside theirs, measured separately for MSD’s brands and the rest, and only from quarters already known when the forecast was made. On 152 MSD forecasts the calibration had not seen it holds 80%; the median band widens from 23% to 35% above P10.
- What the model leans on. Share of the final model’s split gain: growth over the last four quarters 43%; growth last quarter 13%; quarters since launch 9%; quarters ahead 7%; average units over the last four quarters 6%. Recent growth counts for more than the level, so a brand that turns sharply is where it misses most.
- Baselines. Last quarter’s value, the same quarter last year, and the recent trend (capped at about 25% a quarter). Same price path for every method, so the comparison is about volume.
- Limits. US Medicaid only, before manufacturer rebates: not MSD revenue, not the GCC. Vaccines are not in this data. The calibrated range is checked on 152 MSD forecasts, so its coverage is known to within a few points, not exactly. The assistant answers only what its tools cover. An independent case study, not produced or endorsed by MSD.