Demand Forecasting and Safety Stock for Ecommerce

Most brands do not stock out because they forecast badly. They stock out because their supplier is late, and nobody built a buffer for that.

Here is the whole thing in two lines. Safety stock is the extra inventory that covers surprises. Reorder point is the stock level that tells you to place the order today.

This guide gives you both formulas and five worked examples in US dollars. It also covers three things most guides skip:

• How to fix your demand data before you forecast, because stockouts corrupt your sales history
• How to check whether your forecast is actually better than doing nothing
• How to price your buffer correctly, since most guides compare the wrong two numbers

You can run all of it in a spreadsheet tonight.

Quick Answer

Safety stock formula (demand and lead time both vary): SS = Z x square root of [ (L x sigma-d squared) + (d squared x sigma-L squared) ]

Reorder point formula: ROP = (d x L) + SS

Where d is average daily demand, sigma-d is the standard deviation of daily demand, L is average lead time in days, sigma-L is the standard deviation of lead time in days, and Z is your service level factor (1.65 for 95%).

Worked answer: take a SKU that sells 60 units a day. Demand swings by 18 units. Lead time is 30 days and swings by 6. At 95% service it needs 616 units of safety stock and a reorder point of 2,416 units.

Key Takeaways

• Forecast first, buffer second. They are two different jobs and most guides blur them.
• Clean your demand data before you model anything. Stockout days make your history lie.
• Lead time variability usually drives more buffer than demand variability. In our example it caused 93% of the total uncertainty.
• The cost of safety stock is its annual carrying cost, not its purchase value. Getting this wrong makes brands under-buffer.
• Never apply one service level across the whole catalog. That is the most common way brands over-invest in inventory.
• Compare your forecast against a dumb baseline. If it does not beat “last week repeats,” it is not earning its keep.

Part 1: Fix Your Demand Data First

This is step zero and almost nobody covers it. Skip it and every formula below produces a confident wrong answer.

Your sales history is not your demand history. Sales are capped by what you had on the shelf. Demand is not.

The Stockout Doom Loop

Here is how brands quietly starve their own bestsellers.

1. You run out of a SKU for six days
2. Those six days record near-zero sales
3. Next month you forecast using that history, so your average drops
4. A lower average means a lower reorder point and a smaller buy
5. You run out again, sooner
6. Repeat

Each cycle your forecast shrinks. The product looks like it is dying. It is not. You are strangling it.

The fix: flag every stockout day and either exclude it or fill it in with your pre-stockout daily average. Most inventory tools log out-of-stock dates. If yours does not, export your inventory snapshot history and mark any day where on-hand hit zero.

Three More Corrections to Make

Subtract returns. If a SKU has a 22% return rate, 100 units shipped is 78 units of real demand. Apparel and footwear brands who skip this over-forecast every season.

Strip one-off promotions. A three-day flash sale at 5x normal volume inflates your standard deviation. That makes you hold buffer you do not need. Cut it, unless you plan to run that promo again on the same schedule.

Split by channel before you average. Blending Shopify and Amazon into one number hides both patterns. More on this in Part 5.

Rule of thumb: if your data cleanup changes average daily demand by more than 10%, everything downstream was wrong. That is common on the first pass.

Part 2: How Do You Forecast Demand for Ecommerce?

Do not use last month’s total divided by 30. That treats a spike and a dead week as the same signal.

Use exponential smoothing. It weights recent sales more heavily without throwing away history, and it fits in one spreadsheet cell.

New forecast = (alpha x recent actual) + ((1 – alpha) x old forecast)

Alpha is between 0 and 1 and controls how fast you react.

Alpha	Use it for
0.1 to 0.2	Stable, evergreen SKUs
0.3 to 0.4	Normal ecommerce SKUs
0.5 or higher	New launches, trend items, short life cycles

Worked example. Old forecast was 56 units per day. Last week actually sold 68 per day. Alpha is 0.3.

New forecast = (0.3 x 68) + (0.7 x 56) New forecast = 20.4 + 39.2 New forecast = 59.6, round to 60 units per day

The forecast moved up but did not jump to 68. One good week does not become your new normal. Update weekly.

How Do You Know If Your Forecast Is Any Good?

Two checks. Run both monthly. This is the part that separates operators from people who bought a tool.

Check 1: MAPE (Mean Absolute Percentage Error)

For each period, take the gap between forecast and actual. Ignore the sign. Divide by actual. Then average across periods.

Say you forecast 60 units a day for four weeks and actually sold 62, 55, 71, and 58.

• Week 1: |62 – 60| / 62 = 3.2%
• Week 2: |55 – 60| / 55 = 9.1%
• Week 3: |71 – 60| / 71 = 15.5%
• Week 4: |58 – 60| / 58 = 3.4%
• MAPE = 7.8%

Rough reading for ecommerce: under 10% is strong, 10% to 20% is workable, over 30% means your buffer is doing all the work and you should focus there instead.

Check 2: Beat the dumb baseline

Take the simplest possible forecast: next week equals last week. Calculate its MAPE too.

If your smoothing model, your ML tool, or your $600-a-month forecasting app does not beat that naive baseline, it is adding cost and no accuracy. This test kills a lot of expensive software subscriptions, which is exactly why vendors rarely mention it.

Also watch bias. MAPE treats over and under forecasting the same. Add up your raw errors with the signs intact. If the total is consistently negative, you are systematically under-forecasting, and that usually traces back to uncleaned stockout days from Part 1.

Part 3: The BUFFER Framework

Most guides hand you a formula and walk away. That is why the formula gets used once and forgotten.

BUFFER is a six-step loop for setting and maintaining stock levels. Run it per SKU, per channel.

Step	What it means	What you produce
B	Baseline demand	Clean average units sold per day
U	Uncertainty	Standard deviation of demand and of lead time
F	Fill rate target	Service level percent and Z score
F	Fulfillment lead time	Real days from PO to sellable
E	Economics check	Annual carrying cost vs cost of a stockout
R	Refresh rhythm	Recalculation cadence

Most teams jump straight to the fourth step, guess a lead time, and skip uncertainty entirely. That is exactly where stockouts come from.

U: Measure Both Kinds of Uncertainty

You need two numbers, and both come from STDEV.S in Google Sheets or Excel.

• sigma-d: standard deviation of daily demand. Use 90 days of cleaned daily sales.
• sigma-L: standard deviation of lead time. Use your last 8 to 12 purchase orders.

Use actual lead times, not promised ones. The gap between those two numbers is where most of your money is hiding.

F: Pick Your Fill Rate Target

Service level is the share of replenishment cycles where you do not run out. Convert it to a Z score from the normal distribution. The NIST Engineering Statistics Handbook covers the underlying statistics if you want the theory.

Service level	Z score	Stock out roughly
90%	1.28	1 in 10 cycles
95%	1.65	1 in 20 cycles
97.5%	1.96	1 in 40 cycles
99%	2.33	1 in 100 cycles

Tier your catalog. Do not use one number.

• A-tier (top 20% of revenue): 97.5% to 99%
• B-tier (next 30%): 95%
• C-tier (bottom 50%): 90%

And adjust by channel. A marketplace stockout costs you ranking, not just the order. Amazon ties in-stock performance to your ability to keep selling. See the Seller Central inventory guidance. A Shopify stockout costs you one order and an email signup.

If you have not tiered your catalog yet, start with SKU management best practices for multi-channel sellers.

F: Nail Down Real Lead Time

Lead time is not production time. Count every day from PO sent to unit sellable:

• Supplier confirmation and payment clearing
• Production
• Quality check
• Freight, ocean or air
• Customs clearance
• Port to warehouse transit
• Receiving and putaway
• Marketplace check-in, if you use FBA or WFS

That last one catches people. Inventory in an Amazon receiving queue is not sellable inventory. If you are weighing that trade-off, see Amazon FBA vs FBM fulfillment.

Part 4: The Formulas and Five Worked Examples

Safety stock (demand and lead time both vary):

SS = Z x square root of [ (L x sigma-d squared) + (d squared x sigma-L squared) ]

Reorder point:

ROP = (d x L) + SS

Example 1: The Core SKU

A matte black ceramic pour-over kettle sold on Shopify and Amazon.

Input	Value
Average daily demand (d)	60 units
Demand standard deviation (sigma-d)	18 units
Average lead time (L)	30 days
Lead time standard deviation (sigma-L)	6 days
Service level	95% (Z = 1.65)
Landed cost	$11.40
Retail price	$39.00
Contribution per unit after fees and shipping	$18.00

Step 1. Demand piece: 30 x (18 x 18) = 30 x 324 = 9,720 Step 2. Lead time piece: (60 x 60) x (6 x 6) = 3,600 x 36 = 129,600 Step 3. Add and take the root: 9,720 + 129,600 = 139,320. Square root = 373 Step 4. SS = 1.65 x 373 = 615.9, round up to 616 units Step 5. ROP = (60 x 30) + 616 = 1,800 + 616 = 2,416 units

When on-hand hits 2,416, place the PO.

Example 2: The Slow Mover

A specialty kitchen tool. Low volume, erratic demand.

• d = 4 units, sigma-d = 6 units, L = 10 days, sigma-L = 1 day, service level 90% (Z = 1.28)

Demand piece: 10 x 36 = 360. Lead time piece: 16 x 1 = 16. SS = 1.28 x square root of 376 = 1.28 x 19.4 = 25 units ROP = (4 x 10) + 25 = 65 units

Note what flipped. Here demand variability dominates, because the standard deviation (6) is larger than the average (4). Erratic beats slow. And the C-tier service level keeps the buffer honest.

Example 3: The New Launch

No sales history, 60 days in. Short domestic-import mix.

• d = 12 units, sigma-d = 14 units, L = 21 days, sigma-L = 5 days, service level 90% (Z = 1.28)

SS = 113 units, ROP = (12 x 21) + 113 = 365 units

Launches get 90%, not 99%, because your inputs are guesses. Use alpha 0.5 smoothing and recalculate weekly instead of monthly.

Example 4: The Seasonal Peak

A gift set that is dead for nine months and critical for three.

• Peak d = 55 units, peak sigma-d = 20 units, L = 30 days, sigma-L = 6 days, service level 97.5% (Z = 1.88)

SS = 654 units, ROP = (55 x 30) + 654 = 2,304 units

Off-season the same SKU might need 30 units of buffer. Same product, 20x difference. This is why one annual number does not work. More on seasonality in Part 6.

Example 5: The Max-Min Method, and Why It Over-Buys

Most guides on this topic teach only this formula:

SS = (Max daily demand x Max lead time) – (Average daily demand x Average lead time)

Run it on our core SKU, using a peak observed day of 110 units and a worst observed lead time of 44 days:

SS = (110 x 44) – (60 x 30) = 4,840 – 1,800 = 3,040 units

Compare that to 616 units from the statistical method. Nearly five times more inventory, or about $27,600 of extra cash sitting in a warehouse.

Why the gap? Max-min guards against your worst sales day and your worst supplier day landing together, every cycle, forever. That pairing is rare. You pay for it every day of the year.

Use max-min only if: you have fewer than 30 data points, a stockout is catastrophic, or you genuinely have no lead time records. Otherwise it is an expensive shortcut. If you arrived here holding this formula because another guide taught it, this is the upgrade.

The Insight Most Guides Get Half Right

Look at the two variance pieces from Example 1 again.

• Demand variability contributed 9,720
• Lead time variability contributed 129,600

Lead time caused 93% of total uncertainty. Demand caused 7%.

Plenty of guides note that lead time variability increases safety stock. Almost none tell you to calculate the split and use it as a diagnostic. That is the difference between a fact and a decision.

Here is the decision. In this SKU, a perfect demand forecast would barely move the buffer. Tightening the supplier moves it enormously.

Drop lead time variability from 6 days to 2 days:

• Lead time piece: 3,600 x 4 = 14,400
• New total: 9,720 + 14,400 = 24,120. Square root = 155
• New safety stock: 1.65 x 155 = 257 units

From 616 units to 257. That is 359 fewer units and $4,093 of cash freed on one SKU, with no drop in service level.

You did not forecast better. You made your supplier predictable.

Run this split on your top 10 SKUs first. Whichever piece is bigger tells you where to spend the next month:

If the bigger piece is	Work on
Lead time	Supplier terms, freight lanes, backup vendors
Demand	Forecast model, promo planning, data cleanup

How to shrink lead time variability:

• Lock delivery windows into the PO with agreed penalties
• Ask for weekly production status, not one ship confirmation
• Book freight earlier, use one primary lane instead of chasing spot rates
• Split large POs into two staggered shipments
• Qualify a backup supplier before you need one

Our guide to strategic sourcing methods that save ecommerce costs goes deeper on supplier terms.

E: The Economics Check Most Guides Get Wrong

This is worth slowing down for, because the standard version of this comparison is mathematically wrong and it makes brands under-buffer.

The common framing: “616 units of safety stock costs $7,022, and a stockout costs $5,400, so the buffer is barely worth it.”

That comparison is invalid. You are putting a one-time asset purchase against a recurring loss. The $7,022 is not spent, it is converted into inventory you will still sell.

The real yearly cost of safety stock is its carrying cost. That means storage, insurance, shrinkage, spoilage, and the return you gave up by tying that cash down. For most brands it runs 20% to 30% a year.

At 25%:

• Safety stock value: 616 x $11.40 = $7,022
• Annual carrying cost: $7,022 x 0.25 = $1,756 per year

Now compare properly.

• A five-day stockout costs 300 units x $18 contribution = $5,400 per event
• Plus rank recovery, plus ad spend pushed to an unbuyable page

So $1,756 a year of real cost prevents an event that costs $5,400 each time it happens. Not close.

Now price the upgrade from 95% to 99%.

• Safety stock at Z = 2.33: 870 units
• Extra units: 254
• Extra inventory value: $2,896
• Extra annual carrying cost: $724

$724 a year to move from stocking out once in 20 cycles to once in 100. On an A-tier SKU with live ad spend, that is one of the cheapest decisions on your P&L. Framed as “$2,896 of extra inventory,” most founders say no. Framed correctly, it is obvious.

If you have not calculated your own contribution number, start with contribution margin for ecommerce.

Part 5: How Do You Forecast Across Multiple Channels?

The mistake: blending Shopify, Amazon, and Walmart into one total, forecasting that, and wondering why the numbers feel off.

Forecast each channel separately. Then pool the inventory decision.

Why separate forecasts:

• Different demand shapes. Amazon spikes around marketplace deal events. Shopify spikes around your email sends and paid pushes. A blended average erases both.
• Different lead times. FBA needs check-in days. Your own 3PL does not.
• Different stockout costs. Marketplace outages cost ranking. DTC outages cost one order.
• Different control. You own your promo calendar. You do not own theirs.

The Pooling Advantage

Here is the part nobody covers, and it is worth real money.

Pooled inventory needs less total buffer than split inventory. Hold three separate buffers at three locations and you carry three separate cushions, each sized for its own worst case. Hold one central buffer and the highs in one channel offset the lows in another.

Rough guide: pulling buffer from three locations into one can cut total safety stock a lot at the same service level. Variability does not add in a straight line. It adds through square roots.

The trade-off is transit time and labor, which is exactly the decision in 3PL vs in-house vs hybrid fulfillment.

How to run it:

1. Pull daily units per SKU per channel for 90 days
2. Clean each channel separately using Part 1
3. Run exponential smoothing per channel
4. Calculate lead time per channel, including check-in days
5. Calculate safety stock per channel at its own service level
6. Hold buffer centrally where you can, replenish channels weekly

For the systems side, see our roundup of software tools for managing warehouse inventories and the setup detail in Shopify inventory management tools and tips.

Part 6: Seasonality, Peaks, and Promotions

The base formula is static. It produces one number from historical data. Seasonal products break it, because both the average and the variability move.

Pre-Peak Build

Start building 8 to 12 weeks before peak, counting backwards from your full lead time.

1. Find your peak multiplier from last year. If you normally sell 15 a day and sold 55 a day in the four weeks before peak, your multiplier is roughly 3.7x.
2. Recalculate safety stock using peak d and peak sigma-d, not annual averages (see Example 4).
3. Place the PO early enough that stock lands before demand starts, and add days because factories and freight are congested in the same window everyone else is buying.

Post-Peak Drawdown

Deliberately let the buffer thin out. Drop the service level target on seasonal SKUs from 97.5% to 90% or lower for the 4 to 6 weeks after peak. That tells your reorder logic to buy less and lets stock run down through normal sales rather than sitting as dead inventory for nine months.

For hard-dated product like holiday-specific items, set a liquidation date instead of carrying it to next year.

Promotional Spikes

Flash sales, creator posts, and paid pushes are not in your historical sigma-d, because you excluded them in Part 1. Add a one-time buffer on top:

Promotional buffer = (expected daily lift x promotion days) x uncertainty factor

Use 1.2 to 1.5 as the uncertainty factor. A 1.3 factor means you are prepared for the promo to beat your estimate by 30%.

After the promotion, log actual lift versus your estimate. After three or four promos you will have a real lift model instead of a guess.

When Should You Not Add More Safety Stock?

More buffer is not always right. Shrink or skip it when:

• The SKU is being discontinued. Buffer becomes write-off.
• The product is perishable or dated. Expiry cost beats stockout cost.
• Storage fees are high. Long-term marketplace storage fees can erase the margin you were protecting.
• Lead time is very short. If you restock in five days, a 30-day cushion is dead cash.
• It is a slow mover with thin margin. A 99% service level on a SKU selling two a week is a rounding error protected by real money.
• Cash is your binding constraint. If it is inventory versus payroll, cut service levels on C-tier first.
• Your data is dirty. Buffer built on uncleaned stockout data is guesswork with a formula wrapped around it. Fix Part 1 first.

The point of safety stock is to protect profit. When it stops doing that, cut it.

R: Refresh Rhythm

Inputs drift. Set a cadence and hold it.

Frequency	Task
After every PO	Log actual lead time in days
Weekly	Update demand forecast, flag stockout days
Monthly	Recalculate SS and ROP for A-tier, check MAPE and bias
Quarterly	Full catalog recalculation, update sigma-L, re-tier by revenue
Pre-season	Rebuild seasonal SKUs on peak inputs

Logging actual lead time after every PO is the cheapest habit on this list and it powers everything else.

What Is the Fastest Way to Start?

One action today:

1. Pick your top 5 SKUs by revenue
2. Export 90 days of daily units sold
3. Flag and fix stockout days, subtract returns
4. Run AVERAGE and STDEV.S on the cleaned column
5. Pull your last 8 POs, calculate average and standard deviation of actual lead time
6. Run the formula at Z = 1.65
7. Calculate the variance split and see which side is bigger
8. Compare the result to what you hold today

Most brands find one of two things. Either they are badly under-buffered on their bestseller, or they are sitting on thousands of dollars of dead cash in the long tail. Usually both, in the same catalog.

Frequently Asked Questions

The Bottom Line

Forecasting is not about being right. It is about being ready when you are wrong.

Get four things in place and most of your stockouts disappear:

• Clean demand data with stockout days corrected
• A weekly forecast that reacts without overreacting, checked against a dumb baseline
• Safety stock built from actual lead time records, not supplier promises
• A reorder point your team actually watches, priced on carrying cost

The brands that scale cleanly are not the ones with the fanciest forecast. They are the ones with clean data, tight supplier records, and a rule everyone follows.

Want this running without adding headcount? AcquireX builds dedicated offshore teams that own inventory planning end to end: data cleanup, PO tracking, reorder triggers, and supplier follow-up. Not a tool you have to learn. A team that runs it.

See how our supply chain management and procurement and sourcing teams work, or talk to us about your top 20 SKUs.