Mark stockouts
Sales in a stockout week are treated as a lower bound on demand, not automatically as the full demand.
Zenix Forecast Model
A small, explainable experiment for a real supply-chain problem: when a product runs out, sales can be lower than customer demand.
Research status: this prototype has not been benchmarked against TimesFM 3.0, Chronos-2, or a broad set of baselines. One worked example demonstrates the arithmetic; it does not prove accuracy, novelty, or superiority.
Try the example →THE IDEA · IN FOUR SHORT STEPS
Treat sales carefully when stock runs out.
Compare two periods
Use uncensored history to compare an older demand level with the most recent level.
Adapt cautiously
If the levels differ, blend them. The size of the shift and amount of recent evidence set the blend weight.
Respect the lower bound
Keep the estimate at least as high as the largest recorded stockout sale. If all data is censored, report that demand is not identifiable.
Forecast=max( largest stockout sale, (1 − α) × older level + α × recent level )
Adaptation weight: α is capped at 0.80 and grows with the observed shift and the number of recent uncensored observations. The default shift threshold is 10%. These are prototype choices, not tuned or validated settings.
INTERACTIVE · RUNS IN THIS BROWSER
Try a stockout example
Enter weekly observed sales and the 1-based week numbers where stock ran out. This calculator does not send or save your entries.
ZENIX · NEXT WEEK ESTIMATE22.4units · illustrative point estimate
Naïve last-2-sales average15.0 units
Older uncensored level20.0 units
Recent uncensored level23.0 units
Observed demand lower bound16 units
Recent uncensored demand is 15% above the older level. The model adapts by 80%, producing 22.4 units; the stockout observations set a 16-unit lower bound.
A FICTIONAL WORKED EXAMPLE
Where did 22.4 units come from?
All values below are invented for teaching. The future actual is shown only to calculate a one-point error after the prediction.
| Weeks | Sales | Status | How the model uses them |
|---|---|---|---|
| 1–4 | 20, 21, 19, 20 | Stock available | Older level = 20 units |
| 5–8 | 23, 24, 23, 22 | Stock available | Recent level = 23 units |
| 9–10 | 14, 16 | Stockout | Lower bounds; not treated as exact demand |
| Next week | 24 | Synthetic actual | Held out of the forecast input; used only for this example’s error |
In this one constructed case: 22.4 is 1.6 units below the synthetic actual of 24. The naïve average of the last two reported sales is 15, which is 9 units below 24. This is an illustrative example chosen to show why stockout flags matter—not a fair benchmark, proof of correctness, or evidence that Zenix wins generally.
WHAT THIS PROTOTYPE DOES NOT DO
Useful hypothesis; limited model.
- It is a deterministic one-step heuristic, not a trained AI or a full probabilistic forecasting system.
- It does not model seasonality, promotions, prices, substitutions, backorders, multiple products, or future horizons.
- If too few observations are uncensored, true demand may not be identifiable; the model should report a lower bound rather than invent a point estimate.
- Historical choices may be confounded. This demo makes no causal claim about promotions or replenishment.
- There has been no rolling-origin study, ablation, calibration test, or head-to-head test against TimesFM 3.0, Chronos-2, or statistical baselines.
- Do not use this prototype for purchasing, staffing, financial reporting, or operational commitments.
What would count as evidence? Freeze a leakage-safe test design, compare against suitable naïve and censor-aware baselines on identical forecast origins, report uncertainty and errors by regime, and run ablations. Until that work is done, any performance claim is pending.
BACKGROUND READING · NOT AN ENDORSEMENT
Related forecasting research
These sources establish relevant context and prior work. Zenix has not been independently reviewed for novelty or compared with these models.
- Mersereau (2015), “Demand Estimation from Censored Observations with Inventory Record Inaccuracy,” Manufacturing & Service Operations Management. Journal article on learning demand from censored sales observations.
- Trapero, Holgado de Frutos & Pedregal (2024), “Demand forecasting under lost sales stock policies,” International Journal of Forecasting, 40(3), 1055–1068. Peer-reviewed journal article proposing a Tobit Kalman filter for lost-sales demand forecasting.
- Google Research, TimesFM 3.0 (31 August 2026), official research announcement. Describes a multivariate time-series foundation model and its reported benchmark results.
- Ansari et al. (2025), “Chronos-2: From Univariate to Universal Forecasting,” arXiv preprint. A model report describing zero-shot univariate, multivariate, and covariate-informed forecasting.