Demand Planner — LUNE Workspace

Forecast Comparison — SKU by SKU

The same five SKUs, the same five months, all three forecasts side by side — click any SKU to see how each one got there.

There are three different forecasts on this page, and they do not agree. Approach 1 — Seasonal Baseline is the engine that runs today, and the one behind the Inventory Flow grid: a base level × a seasonal shape × a growth factor, plus committed wholesale. Approach 2 — AI-Tuned Growth is that same engine with one dial — how much growth it may assume — set per product by a panel of four AI agents. Approach 3 — Agent-Built Forecast is a separate forecast built from scratch: four statistical models are backtested per product, the winner becomes the baseline, and a crew of six agents then adjusts it month by month with named, unit-sized adjustments. This pane picks no winner and forecasts nothing itself. Oct 2026 – Feb 2027 is used because it is the only window all three cover.

Approach 1 — Seasonal Baseline

The engine running today · drives the Inventory Flow grid

Fits a base level and a seasonal shape from 57 months of history, multiplies by a growth factor, and adds committed wholesale orders. Its growth factor is capped at 1.0× for any product selling 18 months or more — it will not extrapolate growth. Measured average error runs ~49% one month out to ~64% six months out, and it historically comes in low.

Approach 2 — AI-Tuned Growth

Same engine, one dial set by AI agents

Four agents — lifecycle, seasonality, inventory risk, and an adjudicator — decide only how much growth this product is allowed, from a fixed list. They never write a sales number. Tested on 108 months of real history: 31.2% average error against the engine’s 32.9%, and it halves the tendency to under-call.

Approach 3 — Agent-Built Forecast

A separate forecast, built from scratch

Backtests four statistical models per product on rolling 6-month windows and picks the winner, then six agents — data auditor, seasonality, trend, events, synthesizer, critic — adjust it month by month. Every adjustment is named and sized in units, and the arithmetic is checked in code. It also publishes a low–high range per month.

SKU	Approach 1 Seasonal Baseline	Approach 2 AI-Tuned Growth	Approach 3 Agent-Built	How far apart	Same 5 months last year
5540 Recycled Plastic Dermaplaners 12pc Set - T	526,667	521,571	560,140	7%	535,236	close agreement
6622 Recycled Nylon Medium Hair Elastics 12pc S	757,726	769,596	658,020	17%	409,157	some disagreement
6745 Rice Water Protein Shampoo Bar - Strengthe	203,917	159,092	139,120	47%	360,261	they disagree sharply
7490 Stainless Steel Gua Sha	312,063	281,904	221,000	41%	206,382	they disagree sharply
7568 Satin Heatless Curling Set - Sunset Tie Dy	291,431	276,235	275,451	6%	334,337	close agreement
ALL 5	2,091,804	2,008,398	1,853,731		1,845,373	vs last year 1.13× / 1.09× / 1.00×

They nearly agree on 5540 and 7568, and are far apart on 6745 and 7490. The direction is consistent — Approach 1 highest, Approach 3 lowest, Approach 2 between — implying +13%, +9% and flat against the same five months last year. None of them has been scored on these months, because the months have not happened. Click any SKU below to open its full reasoning.

Actual sales, last 18 months, then the three forecasts

Actual sales Approach 1 — Seasonal Baseline Approach 2 — AI-Tuned Growth Approach 3 — Agent-Built

Forecast	OCT 26	NOV 26	DEC 26	JAN 27	FEB 27	TOTAL
Approach 1 — Seasonal Baseline	110,493	86,904	83,240	126,242	119,788	526,667
Approach 2 — AI-Tuned Growth	109,351	86,353	82,380	124,937	118,550	521,571
Approach 3 — Agent-Built	112,600	131,450	100,750	110,240	105,100	560,140
Approach 3 range	88,000–143,000	85,000–163,000	78,000–126,000	85,000–140,000	83,000–128,000

How each approach got to its number

Approach 1 — Seasonal Baseline · 526,667 units

It multiplies a base level of 112,704 units a month by each month’s seasonal shape, then by a growth factor of 1.00×, and adds committed wholesale.

Month by month, the seasonal shape it applies is:

Oct 26 — season 0.98× → 110,493
Nov 26 — season 0.47×, plus 33,600 committed wholesale → 86,904
Dec 26 — season 0.74× → 83,240
Jan 27 — season 1.12× → 126,242
Feb 27 — season 1.06× → 119,788

Seasonal shape taken from this product’s own history. Growth is capped at 1.0× for products selling 18 months or more, which is why it is exactly 1.00 here.

Approach 2 — AI-Tuned Growth · 521,571 units

Same engine, but the agents set growth to 1.0× — they chose not to change anything.

Lifecycle analyst — reads it as plateau, wants 1.0×
This is a 56-month-old SKU that already peaked in 2024 (1,361,902 units), fell 22% in 2025 (0.777 YoY) and is still running 0.801 against 2025 on a calendar basis in 2026; the apparent 1.23 last-12/prior-12 is recovery off the mid-2025 trough (2025-04 at 58,063), not a new ramp. The most recent trajectory is flat, not compounding: last-6 vs prior-6 is 1.025, and the last twelve months sit in a tight band around ~100-125k/month (2025-10 100,232; 2026-03 107,262; 2026-06 105,344; 2026-08 101,166) with a 12-month mean of ~101k versus a 2024 mean of ~113k. The uncapped fit of 1.153x is reading the trough-to-plateau rebound as trend, and extrapolating it six months forward would compound a recovery that has already stopped. Base level at cutoff (111,540) already sits at the top of the recent operating band, so no growth multiplier is warranted.

Seasonality analyst — reads it as plateau, wants 1.0×
The last-12m gain (1,216,141 vs 988,405 = 1.23x) is driven almost entirely by a single low-season window: Nov 2025 hit 100,505 against a 0.473 November index (deseasonalized ~212k) and Dec 2025 hit 102,160 against 0.739, while the high-index months went the other way — May 2026 delivered only 73,733 against a 1.592 index (deseasonalized ~46k) versus 177,351 in May 2024. That mismatch means the sku-level seasonal shape, though built on 56 selling months and not flagged weak, is stale: deseasonalized months over Mar–Aug 2026 scatter from ~46k to ~142k around a base of 111,540, so the fitted 1.153 growth is partly the residue of a seasonal curve that no longer matches the year. With last6_vs_prior6 at only 1.025 and annual totals falling (1,361,902 → 1,058,296 → 848,012 YTD), there is no sustained ramp to extrapolate.

Inventory risk analyst — reads it as plateau, wants 1.1×
Trailing 12m is 1,216,141 units vs prior 988,405 (1.23x), but that lift is entirely a step-recovery from the 2025 trough: last 6m 615,673 vs prior 6m 600,468 is only 1.025x, and monthly units have oscillated in a flat 100k-126k band for eleven straight months (Oct-25 100,232 through Aug-26 101,166) with no upward drift. The fitted 1.153x uncapped growth is therefore reading the Sep-Dec 2025 level shift, not an ongoing ramp, and the base_level_at_cutoff of 111,540 already sits ~9% above the trailing-6m mean of 102,612, so upside cushion is partly baked in. At ~1.2M units/yr and $5 wholesale this SKU carries roughly $6M of annual exposure, so a wrong growth multiplier is expensive in absolute units either way — but the flat last6 and the soft seasonal window ahead (Nov 0.473, Dec 0.739, Apr 0.734 already behind us) make markdown risk the larger one; the only genuine stockout claim is the 33,600 committed Nov wholesale block, which the engine adds outside the growth term anyway. Against the engine's measured ~19% under-call on established SKUs, a small hedge above 1.0 is warranted, but nothing near the fitted 1.153.

Adjudicator — settled on 1.0×
All three specialists agree on the regime — plateau, not ramp — and the disagreement is only about whether to hedge above 1.0. The lifecycle and seasonality analysts both show the 1.23 last-12/prior-12 is a rebound off the 2025-04 trough (58,063) rather than compounding growth, and last6_vs_prior6 of 1.025 confirms the recent half-year is flat; calendar YoY remains below 1.0 in both 2024->2025 (0.777) and 2025->2026 (0.801). The seasonality analyst's warning is decisive against the inventory analyst's 1.1 hedge: the lift sits in low-index months (Nov 0.473, Dec 0.739) while the highest-index month eroded badly (May 2026 = 73,733 against a 1.592 index versus 177,351 in May 2024), so the fitted 1.153 is partly a stale seasonal curve, not trend. I resolved against the hedge because base_level_at_cutoff (111,540) already sits ~9% above the trailing-6m mean of 102,612, so cushion is already embedded in the base, and the forecast window is weighted toward soft months. The 33,600 committed November wholesale block is added outside the growth term and needs no multiplier to be captured.

Best argument against it: The inventory analyst's point stands that the engine's blanket 1.0x cap under-calls established SKUs by roughly 19%, and eleven months range-bound at 100k-126k after a trough is arguably a stabilized new floor rather than a decline — if Nov/Dec 2026 repeat their 2025 strength against a 0.473/0.739 index, a 1.0 cap plus a stale seasonal curve will under-forecast the back half of the window materially.

Approach 3 — Agent-Built · 560,140 units

Of four models backtested on this product, “seasonal_naive” won with an average error of 42.1%. That became the starting point; agents then adjusted it.

Named adjustments to the starting point:

Oct 26: 100,232 → 112,600 — trend_level_correction +5,768; event_ulta_october_order +3,600; event_amazon_prime_big_deal_days +3,000
Nov 26: 100,505 → 131,450 — trend_level_correction +4,495; seasonal_correction -3,150; event_tjx_q4_committed_po +29,600; event_ulta_november_step_down -2,900; event_nordstrom_rack +2,900
Dec 26: 102,160 → 100,750 — trend_level_correction +2,840; seasonal_correction -3,150; event_ulta_december_below_run_rate -1,100
Jan 27: 125,836 → 110,240 — trend_level_correction -21,836; seasonal_correction +6,240
Feb 27: 106,503 → 105,100 — trend_level_correction -503; seasonal_correction +0; event_ulta_february_below_run_rate -900

Key assumptions

Underlying level is ~106,000/mo (Jan-Aug 2026 = 848,012 over 8 months) held flat across the horizon, not tapered and not the 2025 level the baseline repeats.
The SKU is effectively non-seasonal at the invoiced level; the only lift carried is January at 1.06 (Jan 2026 125,836, Jan 2024 125,560, against Jan 2025's 64,848).
The TJX 33,600 PO (2026-11-08 delivery) invoices in full in November, and ~4,000/mo of TJX is already embedded in the 106,002 base (observable TJX: 0 Jul, 11,400 Aug, 1,560 Sep MTD).
Ulta forward orders are carried close to their ~8,500/mo run-rate: upside months halved on top of a 0.80x haircut, downside months halved, because the workbook missed high in Aug (10,200 vs 8,160) and catastrophically in Sep (20,160 vs 1,200 at 18/30 days).

What to watch

Does the Ulta Sep order (forecast 20,160, only 1,200 invoiced at 18 of 30 days) ship late into October, get cut, or get cancelled? Late shipment is +8,000 to +15,000 on October on top of the +3,600 carried.
Confirm the Ulta October PO (19,680) is live and shipping. Only +3,600 is in the plan; a full ship at the 0.80x rate would add another ~3,600.
Confirm the TJX 33,600 invoices in November, not December. Check invoiced TJX units by 2026-11-20; if under ~20,000, shift the balance into December. November's 163,000 high and December's 126,000 high are mutually exclusive on this PO — do not size safety stock off both; cap the Q4 upper case at ~410,000.
Is the TJX order a one-time set-in or the start of a repeating program, and are Q1 2027 deliveries planned? Main swing on Feb-Mar.

Critic’s verdict: revise

Actual sales — last 18 months

MAR 25	APR 25	MAY 25	JUN 25	JUL 25	AUG 25	SEP 25	OCT 25	NOV 25	DEC 25	JAN 26	FEB 26	MAR 26	APR 26	MAY 26	JUN 26	JUL 26	AUG 26
82,835	58,063	108,130	115,842	96,534	69,618	65,232	100,232	100,505	102,160	125,836	106,503	107,262	104,494	73,733	105,344	123,674	101,166

Actual sales, last 18 months, then the three forecasts

Actual sales Approach 1 — Seasonal Baseline Approach 2 — AI-Tuned Growth Approach 3 — Agent-Built

Forecast	OCT 26	NOV 26	DEC 26	JAN 27	FEB 27	TOTAL
Approach 1 — Seasonal Baseline	170,649	258,200	155,112	94,748	79,017	757,726
Approach 2 — AI-Tuned Growth	169,583	259,219	158,900	98,507	83,387	769,596
Approach 3 — Agent-Built	122,750	180,500	175,715	91,455	87,600	658,020
Approach 3 range	75,000–175,000	85,000–235,000	85,000–245,000	48,000–135,000	45,000–130,000

How each approach got to its number

Approach 1 — Seasonal Baseline · 757,726 units

It multiplies a base level of 101,259 units a month by each month’s seasonal shape, then by a growth factor of 1.00×, and adds committed wholesale.

Month by month, the seasonal shape it applies is:

Oct 26 — season 1.69× → 170,649
Nov 26 — season 1.12×, plus 144,600 committed wholesale → 258,200
Dec 26 — season 1.53× → 155,112
Jan 27 — season 0.93×, plus 1,000 committed wholesale → 94,748
Feb 27 — season 0.77×, plus 1,000 committed wholesale → 79,017

Seasonal shape taken from this product’s own history. Growth is capped at 1.0× for products selling 18 months or more, which is why it is exactly 1.00 here.

Approach 2 — AI-Tuned Growth · 769,596 units

Same engine, but the agents set growth to 1.2×.

Lifecycle analyst — reads it as steady growth, wants 1.5×
This SKU is still in a real, sustained ramp rather than a plateau: last-12 units of 980,671 vs prior-12 of 594,329 is 1.65x, almost identical to the 2024->2025 full-year ratio of 1.60, so the YoY rate has stopped decaying (4.02x -> 1.60x -> 1.65x) and is holding near 1.6x. The headline 2025->2026 ratio of 1.077 is a partial-year artifact — 2026 YTD of 703,537 covers only Jan-Aug, and the same eight months of 2025 total 376,205, i.e. 1.87x. The raw last6/prior6 of 1.197 looks like deceleration but is seasonality-confounded: the prior-6 window (Sep-Feb) carries seasonal indices summing to 6.86 vs 5.14 for the last-6 window (Mar-Aug), and deflating by that 1.34x ratio puts recent half-year growth back at ~1.60x, in line with the full year. Month-level volatility is high (Apr-2025 at 7,218 vs May-2025 at 83,990; Jul-2026 at 133,239 vs Aug-2026 at 61,035), so I discount the 1.965 uncapped fit and recommend 1.5 rather than letting the full fitted growth through.

Seasonality analyst — reads it as steady growth, wants 1.4×
The recent surge sits squarely in this SKU's LOW-index months, which argues it is real growth rather than season: July's index is the lowest of the year at 0.417 yet 2026-07 posted the all-time high of 133,239 units, and August (index 0.600) still beat 2025-08 (61,035 vs 53,702). Comparing identical months removes seasonality entirely: Mar-Aug 2026 = 534,332 vs Mar-Aug 2025 = 274,093, a 1.95x same-season YoY, and the prior half showed 446,339 vs 320,236 = 1.39x — so the headline 2025->2026 ratio of 1.077 is an artifact of an 8-month partial year, not deceleration. The counterweight is that the forecast window is the high-season half (Sep-Feb indices sum 6.861, avg 1.144, vs 0.856 for the Mar-Aug period the trend was fitted on), so the engine already lifts the base 94,193 by ~14% before any growth factor and a high cap compounds into the seasonal peak on top of 144,600 committed wholesale in 2026-11. The sku index is drawn from 56 selling months and is not flagged weak, but it is internally inconsistent at the month level (2025-04 = 7,218 vs 2026-04 = 95,798 against a flat 0.709 index), so I discount the fitted 1.965 rather than pass it through.

Inventory risk analyst — reads it as steady growth, wants 1.4×
Trailing twelve months of 980,671 units versus 594,329 prior is a 1.65x step-up, and the momentum is not a seasonal artifact: the last 6 months (Mar-Aug 2026, seasonal weight 5.14) delivered 534,332 units against 446,339 in the prior 6 months that carried the heavy Oct-Dec peak (weight 6.86), implying roughly 1.6x seasonally-adjusted growth off a base level of 94,193/mo. At this volume the asymmetry is clearly stockout-side: a 1.0x cap projects ~646k organic units plus 146,600 committed wholesale over Sep-Feb, and would under-buy a run-rate that has averaged ~89k/mo for six months, while the engine already carries a ~19% under-call bias on established SKUs. The counterweight is real but modest — Aug 2026 fell to 61,035 and the 2025->2026 ratio prints 1.077 on a partial year — so I would not let the full fitted 1.965x through; 1.4x captures most of the measured ramp while leaving markdown exposure bounded on a black core-basic that carries low obsolescence risk.

Adjudicator — settled on 1.2×
All three specialists agree the ramp is real and I accept that: every month of 2026 beats its 2025 counterpart, last-12 of 980,671 vs prior-12 of 594,329 is 1.65x, and the headline 1.077 for 2025->2026 is genuinely an 8-vs-12-month artifact — so the mature-SKU default of 1.0x is wrong here. Where I part from the panel is magnitude, and I side with the seasonality analyst's counterweight over the lifecycle analyst's 1.5: the growth signal was measured in the Mar-Aug half whose indices sum to only 5.137, so the deseasonalized level is obtained by dividing by small, demonstrably unstable indices (2026-07 prints 319k deseasonalized on a 0.417 index while 2026-08 prints 102k on 0.600) and is then applied to the Sep-Feb half whose indices sum 6.861 — index error and cap error compound in the same direction. The arithmetic also shows the base is already doing the work: base_level 94,193 x 6.861 = ~646k organic plus 146,600 committed is ~793k against 446,339 actual in Sep-2025..Feb-2026, i.e. the engine implies ~1.78x YoY at a 1.0x cap before any growth factor, and the committed Nov order sits on top of a base and seasonal profile that already embed prior years' wholesale lumps. I therefore trim to 1.2 rather than averaging to 1.4; I also discount the inventory analyst's '~19% under-call bias on established SKUs', which is not a figure in this fact sheet, and I note that 2026-08 at 61,035 vs 53,702 is the weakest same-month YoY in the series and the single most recent read. Ambiguity is in the size of the ramp, not its existence, so the rule pushes me to the lower cap.

Best argument against it: Same-season Mar-Aug is 534,332 vs 274,093 (1.75x even after normalizing the anomalous 7,218 in Apr-2025), the deseasonalized half-over-half level has risen 46.7k -> 53.4k -> 65.1k -> 104.0k per index-unit with no sign of decay, and the fitted base of 94,193 sits below the current deseasonalized level of ~104,016 — on that reading 1.2x barely restores the level already observed and a 1.4-1.5x cap is the honest extrapolation, leaving real stockout exposure on a black core-basic with low obsolescence risk.

Approach 3 — Agent-Built · 658,020 units

Of four models backtested on this product, “ensemble_median” won with an average error of 53.9%. That became the starting point; agents then adjusted it.

Named adjustments to the starting point:

Oct 26: 142,712 → 122,750 — Trend + seasonal level rebuild (level cut to 85,000, Oct index cut to 1.15) -44,962; Deferred September wholesale PO invoices in October +25,000
Nov 26: 110,282 → 180,500 — Trend + seasonal level rebuild (level 85,900) -24,382; TJX committed New Item PO, 2026-11-08 +94,600
Dec 26: 182,116 → 175,715 — Trend + seasonal level rebuild (level 86,700) -56,401; TJX committed PO dated 2026-11-29, December delivery +50,000
Jan 27: 103,607 → 91,455 — Trend + seasonal level rebuild (level 87,100) -12,152
Feb 27: 108,070 → 87,600 — Trend + seasonal level rebuild (level 87,600) -20,470

Key assumptions

Underlying deseasonalised level is 85,000/mo in Oct 2026, growing +1%/mo through Dec and +0.5%/mo after (85,000 / 85,900 / 86,700 / 87,100 / 87,600 / 88,000). This is the midpoint of the ex-July 2026 average (81,471/mo) and the full Jan-Aug 2026 average (87,942/mo).
Seasonal indices refit on 2025-2026 only: Oct 1.15, Nov 1.00, Dec 1.45, Jan 1.05, Feb 1.00, Mar 0.80. The statistical indices (Oct 1.54, Jan 0.764, Feb 0.769, Jul 0.388) are rejected as artifacts of the sub-scale 2022-2024 months.
Both TJX Q4 POs invoice in their stated months and in full — 94,600 in Nov, 50,000 in Dec — and are fully incremental, since TJX shipped only ~5,160 units across Jan-Aug 2026.
Roughly 25,000 of the ~38,000-unit September wholesale shortfall invoices in October rather than being cancelled (Sep MTD Wholesale 1,080 + TJX 1,080 vs 35,160 + 5,160 in Aug).

What to watch

Open order / backlog report as of end-September: did the ~38,000 units of missing September wholesale ship, slip to October, or get cancelled? This single fact moves October by ±25,000.
On-order and in-transit receipts scheduled to land before 2026-12-01. Sep inventory 344,398 minus Oct+Nov as forecast leaves ~41,000 for a 175,715-unit December. Without confirmed receipts, December is a wish, not a forecast — the constrained outcome is ~60,000 with the TJX 50,000 slipping to Q1.
Invoice-date convention on the TJX PO dated 2026-11-29 for 50,000 units. Plan Nov+Dec jointly at 356,215 until settled.
Confirmation that the 94,600-unit TJX New Item order (2026-11-08) is still Committed at full quantity — it is 52% of the November forecast.

Critic’s verdict: revise

Actual sales — last 18 months

MAR 25	APR 25	MAY 25	JUN 25	JUL 25	AUG 25	SEP 25	OCT 25	NOV 25	DEC 25	JAN 26	FEB 26	MAR 26	APR 26	MAY 26	JUN 26	JUL 26	AUG 26
35,238	7,218	83,990	58,741	35,204	53,702	37,182	67,266	56,257	116,429	91,204	78,001	61,456	95,798	97,538	85,266	133,239	61,035

Actual sales, last 18 months, then the three forecasts

Actual sales Approach 1 — Seasonal Baseline Approach 2 — AI-Tuned Growth Approach 3 — Agent-Built

Forecast	OCT 26	NOV 26	DEC 26	JAN 27	FEB 27	TOTAL
Approach 1 — Seasonal Baseline	41,725	48,470	45,071	38,901	29,750	203,917
Approach 2 — AI-Tuned Growth	32,207	37,413	36,482	30,027	22,963	159,092
Approach 3 — Agent-Built	30,500	32,635	27,828	25,894	22,263	139,120
Approach 3 range	24,000–38,500	23,500–43,000	19,000–37,000	17,000–36,000	12,000–32,000

How each approach got to its number

Approach 1 — Seasonal Baseline · 203,917 units

It multiplies a base level of 44,663 units a month by each month’s seasonal shape, then by a growth factor of 0.91×, and adds committed wholesale.

Month by month, the seasonal shape it applies is:

Oct 26 — season 1.02× → 41,725
Nov 26 — season 1.24× → 48,470
Dec 26 — season 1.13×, plus 2,650 committed wholesale → 45,071
Jan 27 — season 1.09× → 38,901
Feb 27 — season 0.87× → 29,750

Seasonal shape taken from this product’s own history. Growth is capped at 1.0× for products selling 18 months or more, which is why it is below 1.00 here.

Approach 2 — AI-Tuned Growth · 159,092 units

Same engine, but the agents set growth to 1.0× — they chose not to change anything.

Lifecycle analyst — reads it as decline, wants 1.0×
This SKU is in a clear, deepening decline, not growth. Annual totals peaked in 2024 at 1,155,178 units and fell to 966,038 in 2025 (0.836x), and 2025->2026 is running at 0.352x; last-12 vs prior-12 is 0.639 while last-6 vs prior-6 is 0.495, so the decline is accelerating, not stabilizing. Monthly units have collapsed from 81,272 in 2025-11 to a 25,505-31,159 range in 2026-07/08, roughly a third of the year-ago level, and the uncapped fitted annual growth of 0.587 is itself a contraction the engine will keep in full. With 56 selling months and a base level of 34,475 at cutoff, there is no ramp here to protect.

Seasonality analyst — reads it as decline, wants 1.0×
Seasonality cannot explain this SKU's collapse: the decline runs straight through both low- and high-index months. July carries the highest fitted index of the year (1.194) yet July 2026 billed only 25,505 units — the weakest month in the entire 36-month window and down from 95,100 in July 2025 — while August (index 1.004) came in at 31,159 vs 75,843 a year earlier. The seasonal shape itself is trustworthy (sku-level index source, 56 selling months, model_flagged_weak false), and the fitted amplitude is mild (0.771 to 1.240), so the upcoming high-index window of Nov 1.240 / Dec 1.135 / Jan 1.088 will lift units off the 34,475 base but is nowhere near enough to reverse last6_vs_prior6 of 0.495. With UNCAPPED fitted growth at 0.587 the cap is not the binding constraint anyway; there is no upside to release.

Inventory risk analyst — reads it as decline, wants 1.0×
This SKU is in a sustained, deepening decline: last 6m units of 216,453 are 0.495x the prior 6m (437,390), and last 12m 653,843 is 0.639x prior 12m, with monthly volume falling from ~88k in 2025-10 to 25,505-31,159 in 2026-07/08. The fitted uncapped annual growth of 0.587 is a decline the engine keeps in full, so the growth cap is a non-binding lever here — any cap above 1.0 changes nothing, and the only thing a high cap could do is invite upside the data does not support. At ~30k units/month and falling, the dominant inventory risk flips from stockout to overstock and markdown on a consumable: a six-month buy anchored on base level 34,475 with seasonal indices peaking at 1.24 already implies ~180-200k units plus the 3,738 committed wholesale in 2026-12, which is above the current run rate's trajectory. Even accounting for the engine's ~19% under-call bias on established SKUs, that bias is measured on plateauing SKUs, not on one running at half of last year's rate.

Adjudicator — settled on 1.0×
All three specialists agree on the regime and the cap, so there is no disagreement to split — the only real question is whether any of them overlooked an upside case, and none exists. The lifecycle analyst's decline read is corroborated rather than confounded by seasonality: the seasonality analyst showed the collapse runs straight through the year's highest-index month (July, index 1.194, billed 25,505 vs 95,100 a year earlier), so the upcoming Nov 1.240 / Dec 1.135 high-season window cannot be read as a recovery signal. The inventory analyst's point is decisive on mechanics: UNCAPPED_fitted_annual_growth is 0.587, a contraction the engine passes through in full, so the cap binds nothing at any value and the only effect of raising it would be to license upside the data does not support. The committed wholesale of 3,738 units in 2026-12 is immaterial against a ~30k/month base and does not change the picture. I set the cap at 1.0 with high confidence, consistent with the mature-SKU default and with the instruction to prefer the conservative value.

Best argument against it: The strongest counter is that 2026-08 (31,159) ticked up from 2026-07 (25,505) and sits near the fitted base of 34,475, so the series may be finding a floor just before a genuinely high-index Nov-Jan window; if the fitted 0.587 contraction keeps compounding on top of that, the engine could under-forecast the winter months and risk understocking a consumable with repeat demand — but a cap above 1.0 is not the lever that would fix that, since the cap only limits upside and the fitted growth is already below 1.0.

Approach 3 — Agent-Built · 139,120 units

Of four models backtested on this product, “seasonal_index_trend” won with an average error of 41.3%. That became the starting point; agents then adjusted it.

Named adjustments to the starting point:

Oct 26: 30,884 → 30,500 — Level + seasonal correction to current regime -3,884; Ulta October order peak (delta over embedded Ulta base) +3,500
Nov 26: 36,830 → 32,635 — Level correction -4,549; Seasonal correction (Nov 1.220 -> 1.112) -2,646; Walgreens + Nordstrom Rack November step +1,500; Ulta November order (delta over embedded base) +1,500
Dec 26: 34,364 → 27,828 — Level correction -3,818; Seasonal correction (Dec 1.178 -> 0.95) -5,912; Meijer Jan-2027 new-item load-in (committed) +2,650; Meijer launch endcap (pitched, risk-adjusted) +544
Jan 27: 30,815 → 25,894 — Level correction -3,144; Seasonal correction (Jan 1.089 -> 0.96) -3,277; Ulta January order returns to cadence +1,500
Feb 27: 25,435 → 22,263 — Level correction -2,452; Seasonal correction (Feb 0.923 -> 0.87) -1,320; Ulta February order, half-weighted +600

Key assumptions

The Apr 2026 step-down (Mar 54,105 → Apr 36,690, 0.68x) is a permanent level shift, not an outlier: inventory never went below 148,887 and never hit zero, so it is demand/distribution loss, not supply.
Deseasonalised level is 27,000 as of October, decaying -2%/month — not the -7%/mo Apr-Sep slope, because Jul-Sep deseasonalised is flat-to-rising (23,103 → 28,275 → ~29,835). Each of those three points is individually shaky (Jul sits on the workbook→NetSuite cutover, Aug is a wholesale lump, Sep is an estimate), which is why a -4%/mo case is published alongside.
The September Ulta order is late, not cancelled, and invoices in September. It is counted once — inside the ~29,000 September full-month estimate that sets the level — and nowhere else.
Embedded Ulta in the base is 2,088/month (Jul 936, Aug 3,240); embedded Walgreens+Nordstrom Rack is 1,807/month (Aug, the only month supplied). Every wholesale event is a delta over those figures.

What to watch

Does Ulta invoice by 2026-09-30? Zero through 18 days against a 5,508 call. If nothing lands, the September actual comes in near 25,000, the deseasonalised level drops to ~25,800, and the whole horizon needs cutting by roughly 1,200/month (~7,000 total).
Meijer PO and ship date for the 2,650-unit Jan-2027 load-in. Dec vs Jan invoicing moves 3,194 units between months (Dec 24,634 / Jan 29,088). Also confirm whether the 1,088 endcap converts from Pitched to committed, and whether any Feb-Mar replenishment follows the launch.
October actual vs 30,500. October seasonality is unforecastable (Oct 2024 71,638 was that year's low, Oct 2025 88,438 its high), so this month is the cleanest single read on whether the 27,000 level holds or the -4%/mo path is right.
Monthly DTC + Amazon. Aug 2026 was 15,729; Sep annualises to ~15,580. Below ~14,500 the support level has broken and every month needs re-cutting toward the low bounds.

Critic’s verdict: revise

Actual sales — last 18 months

MAR 25	APR 25	MAY 25	JUN 25	JUL 25	AUG 25	SEP 25	OCT 25	NOV 25	DEC 25	JAN 26	FEB 26	MAR 26	APR 26	MAY 26	JUN 26	JUL 26	AUG 26
81,832	72,060	85,072	77,337	95,100	75,843	77,129	88,438	81,272	66,968	60,823	62,760	54,105	36,690	34,121	34,873	25,505	31,159

Actual sales, last 18 months, then the three forecasts

Actual sales Approach 1 — Seasonal Baseline Approach 2 — AI-Tuned Growth Approach 3 — Agent-Built

Forecast	OCT 26	NOV 26	DEC 26	JAN 27	FEB 27	TOTAL
Approach 1 — Seasonal Baseline	52,633	70,242	69,074	48,893	71,221	312,063
Approach 2 — AI-Tuned Growth	46,775	62,922	62,370	44,499	65,338	281,904
Approach 3 — Agent-Built	37,000	48,000	71,000	29,000	36,000	221,000
Approach 3 range	29,000–47,000	36,000–62,000	48,000–96,000	21,000–39,000	25,000–48,000

How each approach got to its number

Approach 1 — Seasonal Baseline · 312,063 units

It multiplies a base level of 53,250 units a month by each month’s seasonal shape, then by a growth factor of 1.00×, and adds committed wholesale.

Month by month, the seasonal shape it applies is:

Oct 26 — season 0.99× → 52,633
Nov 26 — season 1.32× → 70,242
Dec 26 — season 1.30× → 69,074
Jan 27 — season 0.92× → 48,893
Feb 27 — season 1.34× → 71,221

Seasonal shape taken from this product’s own history. Growth is capped at 1.0× for products selling 18 months or more, which is why it is exactly 1.00 here.

Approach 2 — AI-Tuned Growth · 281,904 units

Same engine, but the agents set growth to 1.1×.

Lifecycle analyst — reads it as decelerating growth, wants 1.1×
This SKU had a genuine step-change ramp starting 2024-09 (monthly volumes jumped from ~6k to ~19-25k) and compounded at 3.235x then 3.17x YoY through 2025, but that ramp has clearly rolled over: last-12 vs prior-12 is 1.547 while last-6 vs prior-6 is 0.949, meaning all the remaining growth sits in the older half of the year. The most recent six months decline monotonically (49,004 → 43,019 → 40,406 → 33,748 → 30,552 → 27,746), and even after dividing out the SKU's own low-season indices (Jun 0.685, Jul 0.601, Aug 0.732) the deseasonalized level is flat at roughly 38-51k with no upward drift, versus a fitted base of 45,510. With 56 selling months and 2026 YTD only 1.32x 2025 YTD (down from 3.17x), the fitted 1.589x is extrapolating a ramp that has already stopped compounding. This is deceleration into a plateau at a much higher level, not a continuing ramp — but it is not yet a decline, since every recent month is still above its 2025 counterpart (e.g. 2026-06 33,748 vs 2025-06 24,241 = 1.39x).

Seasonality analyst — reads it as decelerating growth, wants 1.2×
The headline softening (last6/prior6 = 0.949) is a calendar artifact, not a turn: the prior-6 window (Sep 2025–Feb 2026) contains the three highest-index months (Nov 1.319, Dec 1.297, Feb 1.337) while the last-6 window is dominated by the trough (Jun 0.685, Jul 0.601, Aug 0.732). On a season-neutral like-for-like basis Mar–Aug 2026 (224,475) vs Mar–Aug 2025 (148,703) = 1.51x, and Sep–Feb 2025/26 vs 2024/25 = 1.58x, so the fitted 1.589 is not merely reading season as growth — the ramp is real. But deseasonalized monthly level inside the last six months is flat and noisy (Mar ~52.1k, Apr ~41.8k, May ~35.4k, Jun ~49.3k, Jul ~50.8k, Aug ~37.9k, mean ~44.5k), essentially equal to the fitted base of 45,510, meaning the base already embeds the step-up and there is no upward drift left to extrapolate. The forecast window Sep–Feb averages an index of ~1.145 (four of six months at or above 0.988, two above 1.29), so any cap multiplies straight into peak season and asymmetrically amplifies over-buy risk.

Inventory risk analyst — reads it as decelerating growth, wants 1.2×
Raw sequential units have fallen every month since March 2026 (49,004 → 43,019 → 40,406 → 33,748 → 30,552 → 27,746) and last6/prior6 is 0.949, but deseasonalising with the SKU's own indices (Jun 0.685, Jul 0.601, Aug 0.732) puts the recent run at roughly 38k–51k/month against a fitted base of 45,510 — a flat-to-softening level, not a continuing ramp, even though Mar–Aug 2026 (224,475) is still 1.51x Mar–Aug 2025 (148,703). At ~28–30k units/month and $6.50 wholesale this is a ~$180–200k/month line, so a stockout in the Nov/Dec window (indices 1.319 and 1.297) is the more expensive error, and the engine's ~19% under-call bias compounds that. But note the cap-1.0 path alone already buys 45,510 x 6.868 ≈ 312k units for Sep–Feb versus 236,618 actual in the year-ago window (+32%), so a modest cap above 1.0 covers the under-call without extrapolating a growth rate the last two quarters no longer support.

Adjudicator — settled on 1.1×
All three specialists agree on the regime (decelerating growth into a plateau at a much higher level), so the only disagreement is 1.1 vs 1.2, and the deciding question is whether any residual upward drift remains to extrapolate. The seasonality analyst correctly demolishes the naive last6/prior6 = 0.949 reading as a calendar artifact and shows the ramp was real (Mar–Aug 2026 224,475 vs 148,703 = 1.51x), but her own deseasonalized series settles the matter the other way: the last six months sit flat and noisy at ~44.5k/month against a fitted base of 45,510, meaning the step-up is already fully inside the base and the fitted 1.589 would double-count it. I resolve toward the lifecycle analyst's 1.1 rather than 1.2 because the forecast window is peak season (Sep–Feb averages ~1.145, with Nov 1.319 and Dec 1.297), so cap error is multiplied where it is most expensive, and because the inventory analyst's own arithmetic shows the cap-1.0 path already buys ~312k units for Sep–Feb versus 236,618 in the year-ago window (+32%) — the under-call is largely covered before any cap is applied. The Dec 2025 outlier (73,617, deseasonalizing to ~56.8k against ~32k neighbours) is a further reason not to let a higher cap compound into the same months. 1.1 keeps a small allowance for the fact that every recent month still runs above its 2025 counterpart, without extrapolating a compounding rate the last two quarters no longer support.

Best argument against it: The level genuinely is ~1.5x year-ago on a season-neutral basis and six consecutive sequential declines are exactly what this SKU's own trough indices (Jun 0.685, Jul 0.601, Aug 0.732) would produce even on a flat-to-rising level — so if any of that 1.5x momentum persists into Sep–Feb, 1.1 under-buys the two largest months of the year on a ~$180–200k/month line, and a Nov/Dec stockout is the costlier error.

Approach 3 — Agent-Built · 221,000 units

Of four models backtested on this product, “ensemble_median” won with an average error of 40.8%. That became the starting point; agents then adjusted it.

Named adjustments to the starting point:

Oct 26: 50,198 → 37,000 — Level correction — ensemble carries too much YoY growth -17,198; Amazon Prime Big Deal Days + early Q4 pipeline fill +4,000
Nov 26: 65,663 → 48,000 — Level correction — ensemble carries too much YoY growth -32,163; Black Friday / Cyber Monday (Amazon-led) + holiday retail load-in +13,800; Nordstrom Rack Q4 buy above run-rate +700
Dec 26: 76,149 → 71,000 — Level correction — ensemble carries too much YoY growth -42,649; December gifting peak + Q4 retail load-in (capped) +36,800; Nordstrom Rack December buy above run-rate +700
Jan 27: 40,506 → 29,000 — Level correction — post-holiday underlying level -10,806; Nordstrom Rack reverts to base -700
Feb 27: 51,847 → 36,000 — Level correction — ensemble leans on the Feb-2025 load-in lump -19,847; Valentine's gifting / Q1 spa-tool demand +4,000

Key assumptions

September 2026 closes at 29,000-30,000 (0.96-0.99x Sep-2025's 30,236, 1.05-1.08x Aug-2026). MTD 12,657 in 18/30 days is 0.76x August against a company-wide 0.68x aggregate / 0.69x median, so the SKU is running 1.12x market pace; Amazon (0.77x) and wholesale (0.75x) soften proportionally, which is an invoicing-lag signature, not a lost account.
YoY growth continues decelerating from +30% (Jun-Aug) to roughly +13-16% on ordinary Q4 months and +6-15% in Q1-2027. No demand break is assumed; the plateau that peaked at 49,004 in Mar-2026 simply does not repeat.
December 2025's 73,617 was part genuine gifting and part one-time Q4 retail load-in, so the Dec/Nov multiplier is capped at 1.48x versus 1.76x observed. Channel split starts only 2026-07, so this cannot be verified either way — it is the single largest judgment in the plan.
Amazon stays at 85-90% of units with normal Q4 deal participation (PBDD +4,000, BFCM +13,800). Wholesale is a rounding error: the Nordstrom Rack forecast (~1,100/mo) is already inside actuals, with only +700 incremental in Nov and Dec.

What to watch

September 2026 final invoiced units, due early October. Above 32,000 lifts the whole horizon ~8-10%; below 27,000 means the Amazon decline is structural and Oct-Nov should come down to ~33,000 / ~43,000.
Amazon deal acceptance and discount depth for Prime Big Deal Days (Oct) and BFCM Lightning/Best Deals. With ~85% of units on Amazon this is the biggest swing in Oct-Dec and it is not yet in any memo.
Holiday PO dates and quantities from Ulta, Target, Walmart, TJX and Walgreens — none are in the wholesale forecast, yet they likely drove part of Dec-2025's 73,617. Late POs shift units from December into January.
Inventory position versus the ~2.8 months of Q4 cover: confirm the December buy is placed by about 2026-10-01 given the 60-day lead time, sized to 71,000 with exposure to 85,000.

Critic’s verdict: revise

Actual sales — last 18 months

MAR 25	APR 25	MAY 25	JUN 25	JUL 25	AUG 25	SEP 25	OCT 25	NOV 25	DEC 25	JAN 26	FEB 26	MAR 26	APR 26	MAY 26	JUN 26	JUL 26	AUG 26
32,650	15,859	29,582	24,241	28,170	18,201	30,236	31,812	41,824	73,617	25,147	33,982	49,004	43,019	40,406	33,748	30,552	27,746

Actual sales, last 18 months, then the three forecasts

Actual sales Approach 1 — Seasonal Baseline Approach 2 — AI-Tuned Growth Approach 3 — Agent-Built

Forecast	OCT 26	NOV 26	DEC 26	JAN 27	FEB 27	TOTAL
Approach 1 — Seasonal Baseline	53,655	72,712	90,099	42,277	32,688	291,431
Approach 2 — AI-Tuned Growth	50,858	68,920	85,401	40,072	30,984	276,235
Approach 3 — Agent-Built	53,101	66,853	87,246	35,742	32,509	275,451
Approach 3 range	44,000–62,500	53,500–80,500	66,000–105,000	26,500–45,000	22,500–42,500

How each approach got to its number

Approach 1 — Seasonal Baseline · 291,431 units

It multiplies a base level of 46,788 units a month by each month’s seasonal shape, then by a growth factor of 0.96×, and adds committed wholesale.

Month by month, the seasonal shape it applies is:

Oct 26 — season 1.20× → 53,655
Nov 26 — season 1.66× → 72,712
Dec 26 — season 2.10× → 90,099
Jan 27 — season 1.01× → 42,277
Feb 27 — season 0.80× → 32,688

Seasonal shape taken from this product’s own history. Growth is capped at 1.0× for products selling 18 months or more, which is why it is below 1.00 here.

Approach 2 — AI-Tuned Growth · 276,235 units

Same engine, but the agents set growth to 1.0× — they chose not to change anything.

Lifecycle analyst — reads it as decline, wants 1.0×
This SKU is in a sustained, multi-year decline: yearly totals fall 1,191,736 (2023) → 1,047,544 (2024) → 643,945 (2025), with YoY ratios decaying 0.879 → 0.615 → 0.438. The decline is accelerating rather than stabilizing — last-6 vs prior-6 is 0.529, materially worse than last-12 vs prior-12 at 0.758. Month-level data confirms it: the 2025 holiday peak of 106,007 in Dec 2025 was well below Dec 2024's 146,357, and recent trough months (26,675 in Apr 2026, 28,168 in Jul 2026) sit below their year-ago counterparts (45,630 and 35,754). There is no growth to cap here; the uncapped fitted annual growth of 0.767 is itself a decline the engine will keep in full, and no committed wholesale offsets it.

Seasonality analyst — reads it as decline, wants 1.0×
The seasonal shape here is unusually trustworthy: the index is sku-sourced over 56 selling months with four full Q4 cycles, and the fitted indices (Oct 1.199, Nov 1.661, Dec 2.104 vs a Jun trough of 0.584) match the raw series closely — Dec 2023 215,300, Dec 2024 146,357, Dec 2025 106,007. Almost none of the recent movement is growth: the Jul→Aug 2026 uptick (28,168 → 39,676) sits in months whose indices barely move (0.793 → 0.828), and the headline last6/prior6 of 0.529 is exaggerated because the prior window contained the Nov–Dec peak. Stripping season out, the clean deseasonalized signal is last12/prior12 = 0.758 and Aug-on-Aug 39,676 vs 44,259 (0.90), both consistent with the uncapped fitted annual growth of 0.767. The forecast window Sep–Feb is peak-weighted, so the seasonal indices — not any growth assumption — will do the work, and there is no committed wholesale to offset.

Inventory risk analyst — reads it as decline, wants 1.0×
Every horizon points down: 2025->2026 YoY is 0.438, last-12m/prior-12m is 0.758 (579,341 vs 763,863), and last-6m/prior-6m is 0.529 (200,405 vs 378,936), with the fitted uncapped annual growth at 0.767 — a decline the engine keeps in full, so the growth cap has no effect on this SKU's number at any allowed value. The only mildly constructive signal is that the most recent three months have narrowed the gap (Jun-Aug 2026 94,152 vs Jun-Aug 2025 109,613, about -14%, versus -47% across the prior six months), which argues the fall is decelerating rather than reversing. At ~33k-40k units/month and a Sep-Feb window carrying the peak seasonal indices (Nov 1.661, Dec 2.104), the dollar exposure here is concentrated in Q4 units, and on a four-year downtrend the expensive error is overstock and post-holiday markdown, not stockout. The engine's ~19% under-call bias on established SKUs does not offset a 24% twelve-month and 47% six-month volume loss.

Adjudicator — settled on 1.0×
All three specialists agree on regime and cap, so there is no substantive disagreement to split — the only nuance is emphasis. The lifecycle analyst reads the steepening (last6/prior6 = 0.529 vs last12/prior12 = 0.758) as an accelerating decline, while the inventory analyst notes the most recent three months narrowed the gap (Jun–Aug 2026 94,152 vs 109,613, about -14%), suggesting deceleration; the seasonality analyst resolves this cleanly by showing the 0.529 is inflated because the prior window carried the Nov–Dec peak, so the trustworthy deseasonalized read is ~0.76, matching the uncapped fitted growth of 0.767. That reconciliation means the decline is real and steady rather than accelerating, but it is still a decline — Aug-on-Aug 39,676 vs 44,259 (0.90) and four straight down years confirm there is no growth signal to protect. Because the fitted annual growth is below 1.0 and the engine keeps declines in full, the cap is non-binding at any allowed value, so the honest choice is the one that asserts no growth. The Sep–Feb window is peak-weighted (Nov 1.661, Dec 2.104), which concentrates dollar exposure in Q4 and makes overstock the expensive error on a four-year downtrend.

Best argument against it: The last three months are the least-bad stretch in over a year (-14% YoY versus -47% across the prior six), and the Jul→Aug 2026 rise from 28,168 to 39,676 came in months whose seasonal indices barely moved — if that stabilization is the start of a floor, a 1.0 cap plus a 0.767 fitted decline will under-forecast the Q4 peak, which is exactly where the units are.

Approach 3 — Agent-Built · 275,451 units

Of four models backtested on this product, “ensemble_median” won with an average error of 24.8%. That became the starting point; agents then adjusted it.

Named adjustments to the starting point:

Oct 26: 54,801 → 53,101 — Model bias correction (partial) -2,000; Ulta Q4 replenishment buy +2,000; September timing risk (net) -500; Oct/Nov split correction (block held) -1,200
Nov 26: 73,053 → 66,853 — Model bias correction (full) -6,400; Ulta November step-down -1,500; Nordstrom Rack November peak +500; Oct/Nov split correction (block held) +1,200
Dec 26: 93,946 → 87,246 — Model bias correction (partial) -4,700; Ulta absent from December -2,000
Jan 27: 38,242 → 35,742 — Model bias correction (token) -1,000; January Ulta reset below run-rate -1,500
Feb 27: 31,509 → 32,509 — Tie to trend level x seasonal index +1,000

Key assumptions

September 2026 closes near 33,000 (22,423 MTD / the 0.68 company-wide MTD ratio), not the 37,372 naive run-rate. Every month below is anchored to that. Trigger: a close under 31,000 means the Oct-Nov level is ~5% high; over 36,000 means it is ~5% low.
The Sep Ulta shortfall (1,232 MTD vs 5,376 in Aug, 0.23x) is order timing, not lost shelf space. I carry only a -500 risk-weight in October and assume no permanent level step; Amazon at 0.64x Aug is in line with the 0.68x market.
Wholesale plan realizes at ~0.58x, so Ulta ships ~5,500 in Oct, ~2,100 Nov, ~1,600 Dec, ~1,750 Jan — below the ~3,600/mo run-rate in every month except October.
Apr 2026 (26,675), Jul 2026 (28,168) and Aug 2026 (39,676) are wholesale timing artifacts, not level changes; inventory never fell below 112,682, so no dip is supply-censored. Base level ~43,000/mo deseasonalised, eroding ~0.6%/mo.

What to watch

Final September actual vs the 33,000 anchor, and whether Ulta invoices catch up in the last 12 days. This is the single check that validates or breaks October.
Whether the Sep Ulta miss (1,232 vs plan 6,272) was a slipped PO or a cancellation. Cancellation signals a ~4,000-5,000/mo level drop across the whole horizon.
Inbound PO / receipt schedule for Oct-Dec — the biggest single risk in this plan. ~97,000 units needed before mid-December; without it invoiced December prints far below 66,000 with no change in consumer demand.
Ulta's actual October ship vs the 9,520 plan. This recalibrates the 0.58x realization factor used in every month here.

Critic’s verdict: revise

Actual sales — last 18 months

MAR 25	APR 25	MAY 25	JUN 25	JUL 25	AUG 25	SEP 25	OCT 25	NOV 25	DEC 25	JAN 26	FEB 26	MAR 26	APR 26	MAY 26	JUN 26	JUL 26	AUG 26
50,735	45,630	41,366	29,600	35,754	44,259	44,599	66,087	80,711	106,007	43,016	38,516	39,129	26,675	33,449	33,308	28,168	39,676

Approach 1 figures are the live engine run behind Inventory Flow. Approach 2 is that engine fitted to Aug 2026 with the agent-chosen growth setting. Approach 3 is the demand-forecast-pilot run of 19 Sep 2026. 60145 is not shown because Approach 2 does not cover it. These numbers are hardcoded and go stale if any of the three is re-run.

Google Sheets Data

No Google Sheets data synced yet. Click "📊 Sync Sheets" to pull data.

Wholesale Forecast

Jake's periodic wholesale demand forecast — synced from Google Drive or uploaded manually, compared release-over-release for drops, increases, and drift.

High-impact only

Loading Wholesale Variance report…

Inventory Flow

Stock, orders and coverage. Its forecast columns are Approach 1 — Seasonal Baseline, the engine running today.

Hide discontinued

SKU		SALES & STOCK 2.0 M-2 → M+6																	ATTRIBUTES			RUNRATES	INVENTORY			OPEN ORDERS	FORECAST		MOS & SAFETY			ACTION	NOTES
Style	Description	2 Months Back	Previous Month	Current MTD	Projected	M+1	M+2	M+3	M+4	M+5	M+6	Run Rate /mo	On Hand	Net Inv	MOS	MOS (fc)	WOS (fc)	PO rec (fc)	ABC	Long LT	Leadtime (mo)	Avg Cons /mo	Total Inv	Committed	NET Inv	Open PO	Next Wk Fcst	Next 4wk Fcst	MOS OH	MOS Inv+PO	Below Safety	Recommended Order	Comments
Loading inventory flow...

Approach 2 — AI-Tuned Growth

The existing engine (Approach 1), with one dial — how much growth it may assume — set per product by a panel of four AI agents. Tested against real history.

6 SKUs · 3 test dates · 108 months checked · 6.76M units

What the AI actually decides here: one number per product — how much growth the forecast is allowed to assume. It never writes a sales figure; the existing engine still does that. The engine’s default is to assume no growth for any product selling 18 months or more, which is safe on average but badly wrong for a product genuinely taking off. Read the caveats before quoting the headline: the AI beats a simple no-AI rule by 0.16 percentage points over 108 checked months, almost all of it from one product (60145), and repeated identical runs have returned different answers. The forward forecast below has not been checked against anything, because those months have not happened.

How accurate is it? — tested on real history

We froze the data at three past dates, forecast six months forward, and compared against what actually sold. Average error is how far off it was, as a percentage — lower is better. Runs low / high is the direction it tends to miss: a negative number means it under-predicts, which causes stockouts.

Method	Average error	Runs low / high	Months within 20%	Better than basic
Simple average of recent months	36.6%	+7.0%	39%	—	the basic control
The engine as it runs today	32.9%	-11.1%	33%	+10.2%	assumes no product grows
Allow every product 30% growth	32.8%	-5.0%	35%	+10.6%	one setting for everything
A simple rule, no AI	31.4%	-8.6%	33%	+14.4%	the bar the agents have to beat
The AI agents	31.2%	-5.7%	36%	+14.9%	a growth setting chosen per product
Best possible, with hindsight	30.2%	-5.5%	—	+17.7%	picking the right dial after the fact — not achievable

Which products did the AI actually help?

Almost all of the gain is 60145, where the AI correctly spotted real growth the engine refuses to assume. Elsewhere it is close to neutral. That argues for running it only on products that look like they are growing, rather than on everything. Growth allowed is the dial it chose at each of the three test dates — 1.0 means “assume no growth”.

SKU	Product	Error — simple rule	Error — AI agents	AI runs low / high	Growth allowed (Dec 24 / Jun 25 / Dec 25)
60145	Zigzag Headbands 2pc	43.3%	38.5%	-32.7%	1.5 / 1.3 / 1.8
7490	Stainless Steel Gua Sha	31.6%	32.9%	-11.8%	1.3 / 1.2 / 1.4
5540	Dermaplaners 12pc Terracotta	29.3%	30.7%	+1.8%	1.1 / 1.0 / 1.0
6622	Med Hair Elastics 12pc	43.6%	44.2%	-7.2%	1.4 / 1.2 / 1.3
6745	Rice Water Shampoo Bar	17.7%	18.2%	+6.1%	1.0 / 1.0 / 1.0
7568	Satin Heatless Curling Set	22.1%	22.1%	+6.1%	1.0 / 1.0 / 1.0

What we expect to sell — next six months (Sep 2026 – Feb 2027)

Units per month, built from sales history up to Aug 2026 (the last complete month). With AI uses the growth setting the agents chose; Engine today is what the system produces without them. Committed wholesale orders are added on top and shown separately. September is still in progress, so its figure is a whole-month estimate against a part-month actual.

The AI raises the six-month total to 3,087,380 units, 132,249 more than the engine’s 2,955,131 (+4.5%). It changed only three of six products — 60145 and 6622 to 1.2×, 7490 to 1.1× — and left the rest alone, which is the same pattern as the history test: step in where growth looks real, stay out of the way otherwise. Nothing here has been checked against reality, because these months have not happened. Only the history test above has been measured. It still changes no number anywhere else on this page.

SKU		SEP 26	OCT 26	NOV 26	DEC 26	JAN 27	FEB 27	6-MONTH TOTAL
60145 Zigzag Headbands 2pc growth allowed 1.2×	With AI	113,771	157,774	96,907	144,071	150,697	101,844	765,064
	Engine today	109,259	149,232	90,279	132,195	136,389	90,695	708,049
	of which wholesale				30	2,088	406	2,524
	Real multi-year ramp (trailing year 3.18x prior) but deseasonalized demand has been flat at ~108k/month since March, so we override the 1.0 mature cap only modestly — the record August is a seasonal peak, not a new level.
7490 Stainless Steel Gua Sha growth allowed 1.1×	With AI	47,384	46,775	62,922	62,370	44,499	65,338	329,288
	Engine today	45,931	44,982	60,031	59,033	41,786	60,868	312,631
	Real ramp, but it has stopped compounding and is already inside the 45,510 base — allow only 1.1x so we don't multiply a plateau through Nov/Dec peak indices.
6622 Med Hair Elastics 12pc growth allowed 1.2×	With AI	81,969	169,583	259,219	158,900	98,507	83,387	851,565
	Engine today	77,903	158,742	250,274	144,290	88,206	73,573	792,988
	of which wholesale			144,600		1,000	1,000	146,600
	Real ramp confirmed (1.65x last-12 vs prior-12, every 2026 month above 2025), but capped at 1.2x because the fitted base plus peak-season indices plus 146.6k committed wholesale already imply ~1.8x YoY for Sep-Feb, and the growth was measured in low-index months that feed an unstable seasonal profile.
5540 Dermaplaners 12pc Terracotta growth allowed 1.0×	With AI	116,139	109,351	86,353	82,380	124,937	118,550	637,710
	Engine today	116,139	109,351	86,353	82,380	124,937	118,550	637,710
	of which wholesale			33,600				33,600
6745 Rice Water Shampoo Bar growth allowed 1.0×	With AI	31,790	32,207	37,413	36,482	30,027	22,963	190,882
	Engine today	31,790	32,207	37,413	36,482	30,027	22,963	190,882
	of which wholesale				3,738			3,738
7568 Satin Heatless Curling Set growth allowed 1.0×	With AI	36,636	50,858	68,920	85,401	40,072	30,984	312,871
7568 Satin Heatless Curling Set growth allowed 1.0×	Engine today	36,636	50,858	68,920	85,401	40,072	30,984	312,871
ALL 6 — with AI		427,689	566,548	611,734	569,604	488,739	423,066	3,087,380
ALL 6 — engine today		417,658	545,372	593,270	539,781	461,417	397,633	2,955,131

Full month-by-month detail is exported to docs/demand-forecast-approach1-forward.csv.

The history test, month by month (data to Dec 2025, forecasting Jan–Jun 2026)

This is the same six-month exercise, but run on months we already know the answer to. With AI and Engine today are the two forecasts; Actual is what really sold. The small figure is how far the AI forecast was off — within 20%, 20–40%, over 40%. The other two test dates behave the same way and are in the CSV.

SKU		JAN 26	FEB 26	MAR 26	APR 26	MAY 26	JUN 26	6-MONTH TOTAL
60145 Zigzag Headbands 2pc growth allowed 1.8×	With AI	80,954	52,981	76,201	54,432	67,237	63,560	395,365
	Engine today	68,445	42,654	58,415	39,733	46,734	42,066	298,047
	Actual · off by	113,557 -29%	86,952 -39%	112,268 -32%	66,201 -18%	84,435 -20%	82,514 -23%	545,927 -27.6%
7490 Stainless Steel Gua Sha growth allowed 1.4×	With AI	36,834	59,753	33,639	37,859	43,143	26,643	237,871
	Engine today	33,704	53,163	29,101	31,847	35,289	21,190	204,294
	Actual · off by	25,147 +46%	33,982 +76%	49,004 -31%	43,019 -12%	40,406 +7%	33,748 -21%	225,306 +5.6%
6622 Med Hair Elastics 12pc growth allowed 1.3×	With AI	55,767	36,553	45,887	46,940	106,102	78,316	369,565
	Engine today	54,171	34,740	42,667	42,702	94,435	68,198	336,913
	Actual · off by	91,204 -39%	78,001 -53%	61,456 -25%	95,798 -51%	97,538 +9%	85,266 -8%	509,263 -27.4%
5540 Dermaplaners 12pc Terracotta growth allowed 1.0×	With AI	111,761	103,834	102,412	71,773	154,554	133,151	677,485
	Engine today	111,761	103,834	102,412	71,773	154,554	133,151	677,485
	Actual · off by	125,836 -11%	106,503 -3%	107,262 -5%	104,494 -31%	73,733 +110%	105,344 +26%	623,172 +8.7%
6745 Rice Water Shampoo Bar growth allowed 1.0×	With AI	78,321	59,418	65,086	57,284	59,568	49,964	369,641
	Engine today	78,321	59,418	65,086	57,284	59,568	49,964	369,641
	Actual · off by	60,823 +29%	62,760 -5%	54,105 +20%	36,690 +56%	34,121 +75%	34,873 +43%	283,372 +30.4%
7568 Satin Heatless Curling Set growth allowed 1.0×	With AI	52,016	35,256	33,590	27,610	27,080	21,870	197,422
	Engine today	52,016	35,256	33,590	27,610	27,080	21,870	197,422
	Actual · off by	43,016 +21%	38,516 -8%	39,129 -14%	26,675 +4%	33,449 -19%	33,308 -34%	214,093 -7.8%
ALL 6		415,653 -10%	347,795 -14%	356,815 -16%	295,898 -21%	457,684 +26%	373,504 -0%	2,247,349 -6.4%

Over these six months the AI called 2,247,349 units against 2,401,133 actual — -6.4%. The engine alone called 2,083,802 — -13.2%. But look at the individual months. A six-month total this close hides monthly misses from −53% to +110%. The worst is 5540 in May 2026 — 154,554 forecast against 73,733 actual — and that is a seasonal error, not a growth error, so the dial the AI turns could not have fixed it. You buy stock monthly, not in six-month blocks, so the monthly swings are what matter.

The four agents, per product

Lifecycle analyst — is the growth real and continuing, or has it levelled off?

Seasonality analyst — is this “surge” just the busy season?

Inventory risk analyst — is over-ordering or running out the bigger danger here?

Adjudicator — reconciles the three, sets the dial, and must record the best argument against its own answer.

On 60145 the seasonality analyst settled it: November 2025 was the highest month on record (121,723 units) in a month that is normally quiet, and April–May sold 3–7× the year before in the off-season. That is real growth, not seasonality — a distinction a blanket rule cannot make.

What the agents cannot see

No promotion or discount history — dp_promotions holds 0 rows, and the promotion multiplier is 1.0 on 100% of live forecast rows.

No price-change history, no advertising spend, no record of when products were out of stock.

No channel split before 2026-07, so every test date here is whole-SKU only.

The agents are forbidden from explaining demand by marketing, pricing, competitors or stock levels, because none of that data exists here — an explanation asserting it would be invention shaped like insight. The only cause they may name is a committed wholesale order, which is real data. That held across every call.

Approach 3 — Agent-Built Forecast

Six-month forecast on 5 pilot SKUs, built channel by channel — Amazon, DTC, Ulta, TJX and other wholesale forecast separately, then added up. Nothing here feeds the grid above.

Run 2026-09-29 · horizon Oct 26–Mar 27 · invoiced units

What changed (v2, after the 2026-09-24 review). Each channel now has its own forecast. TJX counts only confirmed orders, so its past buys no longer inflate the run-rate. Ulta starts from the wholesale team's plan, with ±10% for months 1–3 and ±40% for months 4–6. Amazon and DTC are corrected for this year's trend, not copied from last year. A crew of AI agents (data auditor, channel-mix analyst, Amazon, DTC and wholesale analysts, synthesizer, critic) adjusts each channel's statistical starting point. Explanations are written only when you click “Explain” or a month.
Channel history before July 2026 is estimated from NetSuite revenue by channel, because unit history by channel doesn't exist before then. Planner inputs (channel discontinuations, rules) live in demand-forecast-pilot/inputs/planner-inputs.json.

SKU	Description	Oct 26	Nov 26	Dec 26	Jan 27	Feb 27	Mar 27	Oct–Mar	vs last year	v1 (Sep 19)
5540	Recycled Plastic Dermaplaners 12pc Set - Terracotta	106,352	146,202	87,086	96,068	88,378	91,466	615,552	−4%	663,170
6622	Recycled Nylon Medium Hair Elastics 12pc Set - Black	80,502	174,106	165,007	86,004	80,503	72,003	658,125	+40%	728,420
7490	Stainless Steel Gua Sha	29,932	36,595	58,221	21,890	28,745	28,272	203,655	−20%	258,500
6745	Rice Water Protein Shampoo Bar - Strengthening	29,117	29,037	26,890	23,727	23,170	24,819	156,760	−62%	163,500
7568	Satin Heatless Curling Set - Sunset Tie Dye	47,010	72,694	90,682	36,984	32,579	35,717	315,666	−15%	308,493

5540 — Recycled Plastic Dermaplaners 12pc Set - Terracotta Oct–Mar 615,552 units, −4% vs last year

Channel share of the last 12 months shown next to each channel. Dominant channel: Amazon 56%. Hover a number for its baseline, adjustments and range.

Channel							Oct–Mar
Amazon 56% · steady	75,600 base 82,655	88,200 base 114,182	63,000 base 81,347	69,300 base 102,954	61,110 base 79,370	64,260 base 83,067	421,470
DTC 1% · steady	2,150 base 600	2,460 base 1,585	1,886 base 1,129	1,435 base 776	1,538 base 868	1,845 base 1,670	11,314
Ulta 9% · planned account	15,744 base 19,680	6,264 base 6,960	7,128 base 7,920	8,748 base 9,720	7,560 base 8,400	7,668 base 8,520	53,112
TJX 4% · po driven	0	33,600	0	0	0	0	33,600
Other wholesale 30% · lumpy wholesale	12,858 base 37,774	15,678 base 37,774	15,072 base 37,774	16,585 base 37,774	18,170 base 37,774	17,693 base 37,774	96,056
↳ Nordstrom Rack (wholesale team plan)	2,964	5,784	5,178	2,960	4,545	4,068	25,499
↳ Accounts with no plan on file (catalog lists: CVS, Kohl's, JCPenney, Harmon, Altar'd State, Lotus Salon)	9,894	9,894	9,894	13,625	13,625	13,625	70,557
All wholesale Ulta + TJX + other	28,602	55,542	22,200	25,333	25,730	25,361	182,768
Total	106,352 92,407–125,821	146,202 129,883–165,157	87,086 76,914–100,636	96,068 80,785–121,749	88,378 73,069–111,009	91,466 76,100–115,004	615,552
Same month last year	100,232	100,505	102,160	125,836	106,503	107,262	642,498