Pharmaceutical Demand Planning: When One Size Fits None

Every quarter, pharmaceutical supply planners face an impossible choice: order too much and watch billions in drugs expire on shelves, or order too little and face emergency shortages during demand surges.

The Problem: The $40 Billion Problem

The challenge isn't forecasting demand in general. It's that every drug's demand behaves completely differently.

• Antivirals like Tamiflu: Swing 40× between winter peaks and near-zero summers. Miss the flu spike and you face emergency backlogs. Overproduce and short-shelf-life inventory expires.
• Antibiotics like Amoxicillin: No seasonal pattern. Demand moves with economic conditions nobody predicted six months ago.
• Chronic medications like Levothyroxine: Steady demand, but tie up massive inventory. Any unexpected shift means either stockouts or overstock.

Traditional forecasting requires building separate models for each drug—taking weeks or months per NDC (National Drug Code). With thousands of products to manage, most companies give up and rely on rules of thumb.

The result: Expired inventory during low seasons. Emergency shipments during surges. Production plans that whipsaw with every unexpected change.

The question: What if one model could forecast all of them, immediately, and tell you which external signals actually matter?

The Solution: Three Drugs, Three Different Stories

We pointed Symnasium agent at three national Medicaid drug-demand series (State Drug Utilization Data), forecasting 2025 quarterly demand (Q1-Q4) from 2018Q1-2024Q4 training data.

Zero per-drug tuning. Seconds to forecast.

The model had access to potential drivers: flu surveillance, RSV activity, price, unemployment, enrollment. Its job: figure out which signals matter for which drugs.

Here's what happened.

Story #1: Oseltamivir (Tamiflu) — The Flu-Driven Surge

The Challenge:

Tamiflu demand is a rollercoaster. Winter flu season brings 40× spikes. By summer, demand approaches zero.

Traditional approach: Cold forecast (no external data) gave 65.94% WAPE—essentially guessing.

The Discovery:

The model tested combinations and found: price + unemployment + enrollment.

The Results:

But here's the key finding: Adding ANY other covariate made it worse.

The model tested price, unemployment, enrollment, RSV—and actively rejected all of them. Flu surveillance alone is the signal. Everything else is noise.

What This Means:

Stop paying for data you don't need. The agent proved that flu surveillance is the ONLY driver worth buying for antiviral forecasting. Don't waste budget on macro data feeds that actively hurt performance.

Tie production directly to flu activity. When flu surveillance shows 0.92 correlation with demand, that's your leading indicator. Forecast the surge before it hits.

Story #2: Amoxicillin — The Macro-Driven Antibiotic

The Challenge:

Amoxicillin isn't seasonal. It doesn't follow flu patterns. So what drives it?

Traditional approach: Cold forecast gave 14.08% WAPE—better than Tamiflu baseline, but still leaving significant error on the table.

The Discovery:

The model tested combinations and found: price + unemployment + enrollment.

The Results:

This proves per-drug driver discovery matters. What moves one drug doesn't move another.

What This Means:

Antibiotic demand follows economics, not disease surveillance. Plan production batches based on macro indicators. The model figured out which ones mattered—automatically—in seconds.

Different drugs need different signals. You can't use the same forecasting approach for your entire catalog. The agent ran the test for each drug and told you which drivers to use.

Story #3: Levothyroxine — The "Boring" Drug That Caught an Anomaly

The Challenge:

Levothyroxine is a chronic thyroid medication. Demand is steady, predictable. Most forecasters set it to historical average and move on.

The Discovery:

Symnasium agent delivered 2.66% WAPE (MASE 0.411) with zero tuning—near the noise floor. Essentially covariate-independent (external signals don't help).

But here's the twist: The model's calibrated bands caught a genuine out-of-90%-band event in 2025Q4 that wider Opus 4.8 intervals missed entirely.

What This Means:

Even "boring" steady-state drugs benefit from foundation models. The tight, honest uncertainty bands give early warnings when something genuinely unusual happens—policy changes, supply disruptions, unexpected demand shifts.

Opus 4.8's loose bands absorb these as noise. You miss the signal.

Story #4: What the Symnasium agent can do and coding agents cannot.

The Challenge:

In 2010-11, Florida cracked down on "pill mills." Almost overnight, opioid shipments dropped as enforcement ramped up.

This is the nightmare scenario: A policy-driven regime shift that breaks every forecast model built on pre-crisis data.

The Test :

We tested Symnasium agent on monthly US opioid shipments (DEA ARCOS dataset, 2006-2014):

• Florida — carries the 2010-11 pill-mill crackdown (the hard case)
• US National — smooth seasonal reference (no regime shift)

The model had access to trends, days in month, unemployment, population—but no special treatment for the changepoint. It had to figure it out.

What we compared it to: Opus 4.8 Ultra Code tasked with building a forecasting model. Our baseline is whatever forecaster Opus 4.8 Ultra Code converges on—for these series, SARIMAX.

The Results:

Florida (with regime shift):

But the real story is robustness. When we look at both models with no covariates:

Why? Opus 4.8 bakes in momentum assumptions into the SARIMAX model that it decided to implement. Foundation models don't. When the trend breaks, the Symnasium agent adapts. Opus 4.8 keeps extrapolating into the void.

The changepoint discovery: The agent cleanly surfaced the changepoint at 2011-04 —the exact mechanism that broke the Opus 4.8 model.

Result: Symnasium agent was ~5× more accurate across the changepoint.

The Tagline: Symnasium Agent Wins on all 3 stores

Antiviral (Oseltamivir): 26.97% WAPE, MASE 0.393 — −59% error vs baseline, flu surveillance alone (r = +0.92)

Antibiotic (Amoxicillin): 6.52% WAPE, MASE 0.209 — −54% error vs baseline, price + unemployment + enrollment

Chronic (Levothyroxine): 2.66% WAPE, MASE 0.411 — near the noise floor, caught a 2025Q4 anomaly wider bands missed

It took seconds instead of weeks-per-NDC model-building cycles.

Regime-shift robustness: Across the 2010-11 Florida opioid crackdown, the agent held to 0.0364 WAPE — ~5× more robust than Opus 4.8 when frozen at zero-shot.

Summary

The Platform Value: Automated Driver Discovery

Symnasium transforms forecasting from a guessing game into a precise strategic instrument by automatically isolating the unique DNA of demand for every product in your catalog.

What this means:

Trust here isn’t a promise that the model is right. It’s the receipt that proves which model is right, for which store, with which inputs, and by exactly how much.

How does Symnasium help you improve your operations and your bottom line