{"id":1297,"date":"2026-03-14T21:22:20","date_gmt":"2026-03-14T21:22:20","guid":{"rendered":"https:\/\/inphronesys.com\/?p=1297"},"modified":"2026-03-14T21:22:20","modified_gmt":"2026-03-14T21:22:20","slug":"when-sop-fails-a-data-driven-survival-guide-for-production-planners","status":"publish","type":"post","link":"https:\/\/inphronesys.com\/?p=1297","title":{"rendered":"When S&OP Fails: A Data-Driven Survival Guide for Production Planners"},"content":{"rendered":"

The Quarterly Fiction Session<\/h2>\n

It’s 9 AM on the first Tuesday of the quarter. Fourteen people are crammed into a conference room that seats ten. The VP of Marketing opens her deck with the words „conservative projection“ \u2014 then reveals a forecast that’s 40% above last year. Nobody blinks. The Sales Director counters with his pipeline, which is somehow both „the strongest ever“ and „subject to timing risk.“ The Operations Director raises a hand to mention that the production line can’t actually run 40% faster without a $2M capital investment, but the conversation has already moved to slide 17.<\/p>\n

Someone suggests splitting the difference between the sales and marketing numbers. The CFO, who has been quietly doing math on his phone, mutters something about inventory carrying costs that nobody acknowledges.<\/p>\n

This isn’t a planning process. It’s a hostage negotiation with spreadsheets.<\/p>\n

If this sounds familiar, you’re not alone. The Sales and Operations Planning (S&OP) process \u2014 that monthly or quarterly cross-functional ritual designed to align demand, supply, and financial plans \u2014 works beautifully in textbooks. In practice, most companies run a version that generates more PowerPoint slides than decisions. The gap between S&OP theory and S&OP reality is where production planners live, quietly trying to figure out how many units to actually build while the executives debate forecast philosophies.<\/p>\n

This post is for those planners. You can’t fix S&OP by yourself \u2014 that’s an organizational problem requiring executive sponsorship, incentive realignment, and often cultural change. But you can<\/em> build your own data-driven toolkit that produces better forecasts than the consensus fiction coming out of those meetings. Here’s how.<\/p>\n

Why S&OP Fails: The Four Root Causes<\/h2>\n

Before building the toolkit, it helps to understand why<\/em> S&OP breaks down. Not to assign blame \u2014 but because the failure mode tells you what your toolkit needs to compensate for.<\/p>\n

Silo Warfare<\/h3>\n

Sales optimizes revenue. Operations optimizes output. Finance optimizes margins. Nobody optimizes the whole. Each function walks into the S&OP meeting with a forecast shaped by their own incentives, and the „consensus“ that emerges is typically either the loudest voice or the last compromise before lunch. The resulting plan isn’t data-driven \u2014 it’s politically mediated.<\/p>\n

No Executive Teeth<\/h3>\n

S&OP was designed as a decision-making<\/em> process. In most organizations, it has devolved into a reporting<\/em> meeting. The executive sponsor shows up, nods through 45 minutes of slides, asks one question about a pet product, and leaves without making a single binding decision about resource allocation, production levels, or inventory policy. Without teeth, S&OP is theater.<\/p>\n

Misaligned Incentives<\/h3>\n

Here’s the dirty secret: Sales teams are typically rewarded for beating<\/em> their forecast, not for forecast accuracy<\/em>. This creates a structural incentive to sandbag \u2014 submit a low forecast, then heroically exceed it. Meanwhile, Marketing inflates numbers to justify campaign budgets, and Operations pads lead times to protect their performance metrics. Everyone is gaming the system rationally. The aggregate forecast is garbage because every input is strategically distorted.<\/p>\n

Perfectionism Paralysis<\/h3>\n

„We can’t do S&OP properly until we have perfect data.“ This is the battle cry of organizations that will never have perfect data. They spend years trying to clean master data, integrate ERP systems, and standardize SKU hierarchies \u2014 all worthy goals \u2014 while making no improvement to their actual planning process. Meanwhile, a planner with 36 months of order history and an R script can outperform the entire S&OP consensus forecast in an afternoon.<\/p>\n

According to Gartner (2024), only 15% of companies report that their S&OP process is effective. Here’s where most organizations actually sit on the maturity curve:<\/p>\n\n\n\n\n\n\n\n\n\n
Level<\/th>\nDescription<\/th>\n% of Companies<\/th>\n<\/tr>\n<\/thead>\n
1 \u2014 Reactive<\/td>\nNo formal process, firefighting<\/td>\n~35%<\/td>\n<\/tr>\n
2 \u2014 Standard<\/td>\nMonthly meetings, no real decisions<\/td>\n~30%<\/td>\n<\/tr>\n
3 \u2014 Advanced<\/td>\nCross-functional alignment, executive-led<\/td>\n~20%<\/td>\n<\/tr>\n
4 \u2014 Proactive<\/td>\nIntegrated planning, scenario modeling<\/td>\n~12%<\/td>\n<\/tr>\n
5 \u2014 Cognitive<\/td>\nAI-augmented, real-time demand sensing<\/td>\n~3%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

If your company is at Level 1 or 2 \u2014 and 65% of companies are \u2014 waiting for organizational S&OP maturity is not a strategy. You need something you can do now<\/em>, with your own data<\/em>, without cross-functional buy-in.<\/p>\n

The Planner’s Toolkit: What You Can Do Today<\/h2>\n

Here’s the good news: the most impactful improvements in forecast accuracy don’t require a functioning S&OP process. They require historical order data, a statistical computing environment, and the willingness to let data do the talking. We’ve organized this into four tiers, ordered by implementation speed.<\/p>\n

Tier 1 \u2014 Today: STL Decomposition (Know Your Demand)<\/h3>\n

Before you can forecast anything, you need to understand your demand patterns. STL decomposition separates your order history into trend, seasonal, and remainder components \u2014 revealing the structure hiding inside what looks like messy data.<\/p>\n

Run STL on your order or shipment data. Classify your SKUs by pattern type: seasonal, trending, intermittent, stable, erratic, or slow-moving. Each pattern demands a different forecasting approach and a different safety stock strategy. You wouldn’t medicate a patient without a diagnosis \u2014 don’t forecast a product without understanding its demand signature.<\/p>\n

STL decomposition will help you understand your demand patterns.<\/p>\n

\"Demand<\/p>\n

The chart above shows six real demand patterns from a mid-size manufacturer’s portfolio. SKU-A has textbook seasonality \u2014 perfect for ETS or ARIMA. SKU-C is intermittent \u2014 zeros punctuated by spikes \u2014 which breaks standard methods entirely and needs Croston’s or Syntetos-Boylan. SKU-E is erratic with no discernible structure. Knowing which<\/em> pattern you’re dealing with is half the battle.<\/p>\n

Tier 2 \u2014 This Week: Statistical Forecasting (Replace the Gut-Feel)<\/h3>\n

Once you’ve classified your SKUs, apply the appropriate statistical forecasting method. For seasonal and trending products (typically 50-70% of a portfolio), ETS and ARIMA auto-selection will produce forecasts that are demonstrably better than both Seasonal Naive and \u2014 this is the key part \u2014 the sales team’s gut-feel forecast.<\/p>\n

The R ecosystem makes this almost embarrassingly easy. The fable<\/code> package’s ETS()<\/code> and ARIMA()<\/code> functions automatically test hundreds of model specifications and select the best one using AICc. No PhD required. You feed in the history, and the algorithm returns a properly calibrated probabilistic forecast with prediction intervals.<\/p>\n

Tier 3 \u2014 This Month: Safety Stock Recalibration (Harvest the Dividend)<\/h3>\n

Better forecasts mean lower forecast error. Lower forecast error means you can carry less safety stock while maintaining the same service level. This is where the CFO’s ears perk up.<\/p>\n

The standard safety stock formula is:<\/p>\n

SS = z \u00d7 RMSE \u00d7 \u221aL<\/strong><\/p>\n

Where z is the service level z-score (1.645 for 95%), RMSE is the root mean square error of the forecast, and L is the lead time in periods. When you cut RMSE by switching from sales gut-feel to statistical forecasting, safety stock drops proportionally. That freed-up working capital is real money \u2014 and it’s the planner’s strongest argument for investing in better forecasting.<\/p>\n

Tier 4 \u2014 This Quarter: Demand Sensing (The Aspiration)<\/h3>\n

The ultimate tier adds leading indicators to your forecasting models: point-of-sale data, Google Trends, weather forecasts, commodity prices, social media sentiment. This is dynamic regression territory \u2014 ARIMA(demand ~ temperature + promotion + price)<\/code> \u2014 and it represents the frontier of what a planner can do independently.<\/p>\n

This tier is aspirational for most teams, but even partial implementation (adding one or two leading indicators to your best SKUs) can produce meaningful accuracy gains. Start with whatever external data you can actually access, not what you wish you had.<\/p>\n

\"Toolkit<\/p>\n

The Showdown: Statistical Forecast vs. Sales Gut-Feel<\/h2>\n

Theory is nice. Let’s look at numbers.<\/p>\n

We set up a controlled comparison using six SKUs from a mid-size manufacturer, each representing a different demand pattern. For each SKU, we have 36 months of actual demand data \u2014 24 months for training and 12 months for testing. We compare two forecasting approaches:<\/p>\n

    \n
  1. Sales Gut-Feel<\/strong>: Simulated consensus forecasts representing what typically emerges from an S&OP process \u2014 biased, smoothed, and lag-adjusted to mimic real-world sales forecasting behavior (sandbag on the upside, overshoot on the downside, miss structural changes).<\/li>\n
  2. Statistical Forecast<\/strong>: Auto-selected ETS or ARIMA model, fit exclusively on historical order data. No human judgment, no adjustments, no politics.<\/li>\n<\/ol>\n

    The metric is MAPE<\/strong> (Mean Absolute Percentage Error) \u2014 the average percentage miss, measured against actual demand over the 12-month test period. Note: the sales forecasts are simulated to reflect common S&OP biases (sandbagging, anchoring, seasonal smoothing) \u2014 not from a specific company, but representative of the patterns we see across manufacturing clients.<\/p>\n\n\n\n\n\n\n\n\n\n\n
    SKU<\/th>\nPattern<\/th>\nSales FC MAPE<\/th>\nStatistical MAPE<\/th>\nWinner<\/th>\nImprovement<\/th>\n<\/tr>\n<\/thead>\n
    SKU-A<\/td>\nSeasonal<\/td>\n27.7%<\/td>\n6.2%<\/td>\nStatistical<\/td>\n77.6%<\/td>\n<\/tr>\n
    SKU-B<\/td>\nTrending<\/td>\n30.8%<\/td>\n4.3%<\/td>\nStatistical<\/td>\n86.0%<\/td>\n<\/tr>\n
    SKU-C<\/td>\nIntermittent<\/td>\n59.8%<\/td>\n48.3%<\/td>\nStatistical<\/td>\n19.2%<\/td>\n<\/tr>\n
    SKU-D<\/td>\nStable<\/td>\n17.1%<\/td>\n9.1%<\/td>\nStatistical<\/td>\n46.8%<\/td>\n<\/tr>\n
    SKU-E<\/td>\nErratic<\/td>\n51.6%<\/td>\n32.6%<\/td>\nStatistical<\/td>\n36.8%<\/td>\n<\/tr>\n
    SKU-F<\/td>\nSlow-moving<\/td>\n104.0%<\/td>\n61.3%<\/td>\nStatistical<\/td>\n41.2%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    Statistical forecasting wins on all six SKUs \u2014 with improvements ranging from 19% to 86%. But here’s what makes this analysis honest: the margin of victory varies enormously. On seasonal and trending demand, the algorithm crushes gut-feel by 77-86%. On intermittent and erratic patterns, the advantage shrinks to 19-37% \u2014 still meaningful, but the absolute error remains high for both approaches. The variation across patterns tells the real story.<\/p>\n

    The Seasonal Blowout: SKU-A<\/h3>\n

    SKU-A has strong, consistent seasonality \u2014 the kind of predictable pattern that statistical models were literally designed for. The ETS(A,N,A) model \u2014 additive errors, no trend, additive seasonality \u2014 captures the summer peak and winter trough with a MAPE of just 6.2%. The sales forecast? A 27.7% miss \u2014 a 77.6% improvement for the algorithm. The classic S&OP failure mode is on full display: the team „agreed“ on an annual average and spread it evenly across months, completely ignoring the seasonal pattern they can see in their own data.<\/p>\n

    \"SKU-A<\/p>\n

    This is the most damning chart in the entire analysis. The blue line (statistical forecast) follows the seasonal curve almost perfectly. The red line (sales consensus) cuts straight through the middle like a highway through a mountain range. Every summer, operations scrambles to produce more than planned. Every winter, excess inventory piles up. The statistical model doesn’t eliminate all error \u2014 but it eliminates the systematic<\/em> error that comes from ignoring seasonality.<\/p>\n

    The Trend Trap: SKU-B<\/h3>\n

    SKU-B shows steady growth with a clear upward trend. The ETS(M,Ad,N) model \u2014 multiplicative errors, damped additive trend, no seasonality \u2014 captures this trajectory and projects it forward, achieving a remarkable 4.3% MAPE. The sales forecast, however, comes in at 30.8% \u2014 anchored to „last year plus a bit,“ a heuristic that systematically underestimates growth and leads to chronic stockouts. The 86% improvement is the largest in our portfolio. It’s the cost of anchoring bias, quantified.<\/p>\n

    \"SKU-B<\/p>\n

    The Harder Wins: SKU-C and SKU-E<\/h3>\n

    Here’s where intellectual honesty matters. SKU-C (intermittent, 59.8% vs. 48.3%) and SKU-E (erratic, 51.6% vs. 32.6%) are statistical wins \u2014 but look at the absolute numbers. Even the winning<\/em> forecast is off by 32-48%. Croston’s method handles the intermittent zeros-and-spikes pattern of SKU-C better than the sales team’s overconfident guessing, and ETS with multiplicative errors does meaningfully better than the sales team’s wild swings on SKU-E \u2014 but neither algorithm can conjure precision where the underlying signal is weak.<\/p>\n

    This isn’t a failure of the method \u2014 it’s information<\/em>. When you know a product’s demand is inherently hard to predict, the correct response isn’t „forecast harder.“ It’s to adjust your safety stock and reorder policies to explicitly account for high uncertainty. Set wider buffers, use min\/max replenishment, and stop pretending the forecast will be accurate. Knowing that your forecast will be 32-48% off is, paradoxically, more useful than pretending it will be 5% off.<\/p>\n

    These are also the SKUs where S&OP adds the most value \u2014 where human intelligence about customer behavior, market shifts, or upcoming orders might narrow that remaining gap in ways no algorithm can extract from history alone.<\/p>\n

    The Portfolio View<\/h3>\n

    Across the full six-SKU portfolio, statistical forecasting wins on every single pattern \u2014 with improvements ranging from 19% to 86%. The biggest gains come from seasonal and trending demand (77-86%), where the algorithm eliminates systematic biases. Even on intermittent and erratic patterns, where absolute errors remain high, the statistical approach delivers 19-37% improvement. The net result: a substantial reduction in portfolio-wide forecast error that translates directly to inventory savings.<\/p>\n

    \"Portfolio-wide<\/p>\n

    The portfolio chart makes the business case visual. For every SKU pattern, the gap between the sales forecast bars (red) and statistical forecast bars (blue) represents pure waste \u2014 overproduction, stockouts, expedited freight, overtime labor, and carrying costs that exist solely because the planning process relies on opinion instead of data. The narrower gaps on intermittent and erratic SKUs are equally telling: they pinpoint where the algorithm’s advantage is real but modest, and where human intelligence about customer behavior or market shifts could add the most incremental value.<\/p>\n

    The Inventory Dividend<\/h2>\n

    Better forecasts translate directly to inventory savings through safety stock reduction. Here’s the math.<\/p>\n

    Safety stock at 95% service level with a 1-month (4-week) lead time:<\/p>\n

    SS = z \u00d7 RMSE \u00d7 \u221aL = 1.645 \u00d7 RMSE \u00d7 \u221a1 = 1.645 \u00d7 RMSE<\/strong><\/p>\n

    Where z = 1.645 (95% service level), RMSE is the root mean square error of the forecast, and L = 1 month. RMSE captures both bias and variance \u2014 so biased sales forecasts get penalized appropriately. When you cut RMSE by switching to statistical forecasting, safety stock drops proportionally. At $25\/unit holding cost, the portfolio-level impact across our six SKUs:<\/p>\n\n\n\n\n\n\n\n
    Metric<\/th>\nSales Forecast<\/th>\nStatistical Forecast<\/th>\nDelta<\/th>\n<\/tr>\n<\/thead>\n
    Total safety stock<\/td>\n2,467 units<\/td>\n1,162 units<\/td>\n-1,305 units<\/td>\n<\/tr>\n
    Holding cost<\/td>\n$61,675<\/td>\n$29,050<\/td>\n-$32,625<\/strong><\/td>\n<\/tr>\n
    Reduction<\/td>\n\u2014<\/td>\n\u2014<\/td>\n52.9%<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    The biggest single wins are SKU-B and SKU-A. SKU-B’s safety stock drops from 646 to 106 units \u2014 an 83.6% reduction \u2014 because the trend-aware ETS model produces dramatically tighter error bands than the anchored sales forecast. SKU-A drops from 473 to 113 units (76.1% reduction) as the seasonal ETS model eliminates the systematic bias from the sales team’s flat-average approach.<\/p>\n

    That’s $32,625 freed from just six SKUs at a $25\/unit holding cost \u2014 with zero service level reduction. Your actual unit costs are almost certainly higher, and your portfolio almost certainly larger. A manufacturer with 200 SKUs averaging $50\/unit holding cost could be looking at a six-figure working capital reduction from forecast improvement alone.<\/p>\n

    \"Safety<\/p>\n

    This is the chart you bring to your next budget meeting. It translates forecast accuracy \u2014 an abstract statistical concept \u2014 into dollars sitting on shelves. Every bar shows money that could be redeployed to growth investments, debt reduction, or simply a healthier balance sheet. The CFO doesn’t care about MAPE. The CFO cares about this.<\/p>\n

    When You Still Need S&OP<\/h2>\n

    Intellectual honesty requires acknowledging what statistical forecasting cannot<\/em> do. Data-driven methods excel at the 80% of your portfolio where history repeats \u2014 but the 20% that requires judgment is exactly where S&OP should focus its limited attention.<\/p>\n

    New Product Launches<\/h3>\n

    No history means no statistical model. When you’re launching a new product, you need analogous product analysis, market research, and judgment. This is a legitimate S&OP discussion topic \u2014 estimating the demand curve for something that doesn’t exist yet.<\/p>\n

    Promotions and Events<\/h3>\n

    A July promotion that’s twice the scale of last year’s won’t be captured by an ARIMA model trained on last year’s data. Promotional lifts need to be overlaid on the statistical baseline, ideally using promotional response curves estimated from past events. This is another place where cross-functional input \u2014 from Marketing and Sales \u2014 genuinely adds value.<\/p>\n

    Strategic Pivots<\/h3>\n

    If you’re entering a new market, exiting a product line, or acquiring a competitor, historical patterns are partly invalidated. Strategic S&OP discussions should focus here \u2014 on the scenarios where the future won’t look like the past.<\/p>\n

    Customer Concentration Risk<\/h3>\n

    If one customer represents 30% of your volume, their internal decisions (changing suppliers, altering order patterns, negotiating new terms) can overwhelm any statistical signal. Regular dialogue with key accounts is irreplaceable \u2014 and it’s an S&OP input, not a statistical one.<\/p>\n

    The point isn’t that S&OP is useless. The point is that S&OP should focus its scarce executive attention on these high-judgment situations \u2014 not on debating whether SKU-A’s July demand will be 1,200 or 1,300 units. Let the algorithm handle the routine. Save human judgment for the exceptional.<\/p>\n

    Interactive Dashboard<\/h2>\n

    Explore the Forecast Accuracy Simulator \u2014 adjust demand patterns, forecast methods, and supply chain parameters to see how statistical forecasting compares to sales gut-feel in real time.<\/p>\n

    \n

    Interactive Dashboard<\/strong><\/p>\n

    Explore the data yourself \u2014 adjust parameters and see the results update in real time.<\/p>\n

    Open Interactive Dashboard \u2192<\/a><\/p>\n<\/div>\n

    Your Next Steps<\/h2>\n

    Don’t wait for your company to fix S&OP. These five actions use only data you already have and tools you can install today.<\/p>\n

      \n
    1. Run STL on your top 20 SKUs and classify demand patterns.<\/strong> Use the R code below to decompose each SKU’s order history. Tag every product as seasonal, trending, intermittent, stable, erratic, or slow-moving. This classification drives every decision downstream \u2014 forecasting method, safety stock policy, and reorder logic. If you haven’t decomposed your demand data, you’re planning blind.<\/li>\n
    2. Benchmark sales forecasts against Seasonal Naive.<\/strong> Compute MASE (Mean Absolute Scaled Error) for whatever forecasts your S&OP process currently produces. If MASE > 1.0, the consensus forecast is losing to „same as last year“<\/em> \u2014 and every dollar of safety stock is being sized against a forecast that adds negative value. This is the number that gets attention in meetings.<\/li>\n
    3. Run a 6-month parallel pilot.<\/strong> Pick 20-30 SKUs spanning different demand patterns. Generate statistical forecasts weekly using ETS\/ARIMA auto-selection. Track both the S&OP consensus forecast and the statistical forecast against actuals. After 6 months, the data will speak for itself \u2014 and you’ll have an evidence base, not an opinion.<\/li>\n
    4. Recalculate safety stock using statistical forecast errors.<\/strong> Replace the S&OP forecast RMSE with the statistical forecast RMSE in your safety stock formula (SS = z \u00d7 RMSE \u00d7 \u221aL). You’ll see immediate inventory reduction opportunities with no service level impact. Quantify the working capital freed in dollars \u2014 that’s your ROI case.<\/li>\n
    5. Bring the data to the S&OP table.<\/strong> Once you have 6 months of parallel-run results, present a one-page comparison: MAPE by SKU, safety stock delta, and working capital impact. Don’t argue about methodology. Show the scoreboard. Let the numbers make the case. The goal isn’t to eliminate S&OP \u2014 it’s to redirect it toward the decisions that actually need human judgment.<\/li>\n<\/ol>\n
      \nShow R Code<\/strong><\/summary>\n
      # =============================================================================\n# When S&OP Fails: Statistical Forecasting vs. Sales Gut-Feel\n# Complete analysis pipeline for 6 SKU demand patterns\n# Run with: Rscript generate_sop_images.R\n# =============================================================================\n\nlibrary(fpp3)\nlibrary(ggplot2)\nlibrary(dplyr)\nlibrary(tidyr)\nlibrary(scales)\nlibrary(patchwork)\n\nset.seed(42)\nimg_dir <- \"Images\"\n\n# =============================================================================\n# COLOR PALETTE & THEME\n# =============================================================================\n\ncol_red    <- \"#e74c3c\"\ncol_blue   <- \"#2980b9\"\ncol_green  <- \"#27ae60\"\ncol_orange <- \"#e67e22\"\ncol_purple <- \"#8b5cf6\"\ncol_dark   <- \"#2c3e50\"\ncol_grey   <- \"#95a5a6\"\n\ntheme_sop <- theme_minimal(base_size = 13) +\n  theme(\n    plot.title    = element_text(face = \"bold\", size = 14),\n    plot.subtitle = element_text(color = \"grey40\", size = 11),\n    panel.grid.minor = element_blank(),\n    legend.position  = \"bottom\"\n  )\n\n# =============================================================================\n# DATA GENERATION \u2014 6 SKUs, 36 months (Jan 2023 \u2013 Dec 2025)\n# Train: 24 months (2023\u20132024), Test: 12 months (2025)\n# =============================================================================\n\nn <- 36\nmonths <- yearmonth(seq(as.Date(\"2023-01-01\"), as.Date(\"2025-12-01\"), by = \"month\"))\n\n# --- SKU-A: Seasonal (~1,000\/month, ~40% summer peak, CV ~15%) ---\nseasonal_idx <- c(-0.32, -0.20, 0.00, 0.20, 0.38, 0.45,\n                  0.42,  0.30, 0.10, -0.10, -0.25, -0.35)\ndemand_a <- round(1000 * (1 + rep(seasonal_idx, 3)) + rnorm(n, 0, 40))\ndemand_a <- pmax(demand_a, 100)\n\n# --- SKU-B: Trending (~800 -> ~1,300, ~+15\/month, mild noise) ---\ndemand_b <- round(800 + (0:35) * 15 + rnorm(n, 0, 40))\ndemand_b <- pmax(demand_b, 200)\n\n# --- SKU-C: Intermittent (~200 when non-zero, ~40% zeros) ---\ndemand_c <- numeric(n)\nfor (i in 1:n) {\n  if (runif(1) > 0.40) {\n    demand_c[i] <- round(rgamma(1, shape = 5, rate = 0.025))\n  }\n}\n\n# --- SKU-D: Stable (~1,500\/month, CV ~8%) ---\ndemand_d <- round(rnorm(n, 1500, 120))\ndemand_d <- pmax(demand_d, 800)\n\n# --- SKU-E: Erratic (~600\/month, CV ~35%) ---\ndemand_e <- round(rnorm(n, 600, 210))\ndemand_e <- pmax(demand_e, 50)\n\n# --- SKU-F: Slow-moving (~50\/month, CV ~45%, many near-zero) ---\ndemand_f <- round(pmax(rnorm(n, 50, 25), 0))\n\n# --- Combine into tsibble ---\nsku_labels <- c(\"SKU-A: Seasonal\", \"SKU-B: Trending\", \"SKU-C: Intermittent\",\n                \"SKU-D: Stable\", \"SKU-E: Erratic\", \"SKU-F: Slow-Moving\")\n\nsku_data <- tibble(\n  month   = rep(months, 6),\n  sku     = factor(rep(c(\"SKU-A\", \"SKU-B\", \"SKU-C\", \"SKU-D\", \"SKU-E\", \"SKU-F\"),\n                       each = n),\n                   levels = c(\"SKU-A\", \"SKU-B\", \"SKU-C\", \"SKU-D\", \"SKU-E\", \"SKU-F\")),\n  demand  = c(demand_a, demand_b, demand_c, demand_d, demand_e, demand_f),\n  label   = factor(rep(sku_labels, each = n), levels = sku_labels)\n)\n\nsku_ts <- sku_data |>\n  select(month, sku, demand) |>\n  as_tsibble(index = month, key = sku)\n\ntrain <- sku_ts |> filter(year(month) <= 2024)\ntest  <- sku_ts |> filter(year(month) == 2025)\n\n# =============================================================================\n# SALES FORECASTS (simulated gut-feel)\n# =============================================================================\n\n# SKU-A: Flat average (ignores seasonality)\nsf_a <- rep(round(mean(train$demand[train$sku == \"SKU-A\"])), 12)\n\n# SKU-B: Year-1 average \u2014 \"last year's budget\" (misses trend)\nsf_b <- rep(round(mean(train$demand[train$sku == \"SKU-B\" &\n                                     year(train$month) == 2023])), 12)\n\n# SKU-C: Overpredicts \u2014 non-zero avg inflated by \"safety\" buffer + noise\nnonzero_c <- train$demand[train$sku == \"SKU-C\" & train$demand > 0]\nset.seed(98)\nsf_c <- round(pmax(20, mean(nonzero_c) * 1.7 * (1 + rnorm(12, 0, 0.35))))\n\n# SKU-D: Close but biased 10% high (optimistic sales team)\nsf_d <- rep(round(mean(train$demand[train$sku == \"SKU-D\"]) * 1.10), 12)\n\n# SKU-E: Anchors high, adjusts wildly (overconfident)\nset.seed(99)\nsf_e <- round(pmax(200, 900 + rnorm(12, 0, 250)))\n\n# SKU-F: Rounds up aggressively\nsf_f <- rep(round(mean(train$demand[train$sku == \"SKU-F\"]) * 1.5), 12)\n\ntest_months <- yearmonth(seq(as.Date(\"2025-01-01\"), as.Date(\"2025-12-01\"),\n                             by = \"month\"))\n\nsales_lookup <- tibble(\n  sku = factor(rep(c(\"SKU-A\", \"SKU-B\", \"SKU-C\", \"SKU-D\", \"SKU-E\", \"SKU-F\"),\n                   each = 12),\n               levels = c(\"SKU-A\", \"SKU-B\", \"SKU-C\", \"SKU-D\", \"SKU-E\", \"SKU-F\")),\n  month = rep(test_months, 6),\n  sales_forecast = c(sf_a, sf_b, sf_c, sf_d, sf_e, sf_f)\n)\n\n# =============================================================================\n# STATISTICAL FORECASTS (fpp3 models)\n# =============================================================================\n\n# SKU-A: Seasonal ETS (additive errors, no trend, additive season)\nets_a_fit <- train |> filter(sku == \"SKU-A\") |>\n  model(stat_fc = ETS(demand ~ error(\"A\") + trend(\"N\") + season(\"A\")))\nets_a_fc  <- ets_a_fit |> forecast(h = 12)\n\n# SKU-B, SKU-D, SKU-E: Auto ETS\nets_bde_fit <- train |> filter(sku %in% c(\"SKU-B\", \"SKU-D\", \"SKU-E\")) |>\n  model(stat_fc = ETS(demand))\nets_bde_fc  <- ets_bde_fit |> forecast(h = 12)\n\n# SKU-C, SKU-F: Croston for intermittent\/slow-moving\ncro_fit <- train |> filter(sku %in% c(\"SKU-C\", \"SKU-F\")) |>\n  model(stat_fc = CROSTON(demand))\ncro_fc  <- cro_fit |> forecast(h = 12)\n\n# Combine all statistical forecasts\nstat_fc_all <- bind_rows(\n  ets_a_fc   |> as_tibble() |> select(sku, month, stat_forecast = .mean),\n  ets_bde_fc |> as_tibble() |> select(sku, month, stat_forecast = .mean),\n  cro_fc     |> as_tibble() |> select(sku, month, stat_forecast = .mean)\n)\n\n# =============================================================================\n# MAPE CALCULATION\n# =============================================================================\n\ntest_df <- test |> as_tibble()\n\ncomparison <- test_df |>\n  left_join(sales_lookup, by = c(\"sku\", \"month\")) |>\n  left_join(stat_fc_all, by = c(\"sku\", \"month\")) |>\n  mutate(\n    sales_ape = ifelse(demand > 0,\n                       abs(demand - sales_forecast) \/ demand * 100, NA),\n    stat_ape  = ifelse(demand > 0,\n                       abs(demand - stat_forecast) \/ demand * 100, NA)\n  )\n\nmape_table <- comparison |>\n  group_by(sku) |>\n  summarise(\n    sales_mape = round(mean(sales_ape, na.rm = TRUE), 1),\n    stat_mape  = round(mean(stat_ape, na.rm = TRUE), 1),\n    .groups = \"drop\"\n  ) |>\n  mutate(improvement_pct = round((1 - stat_mape \/ sales_mape) * 100, 1))\n\ncat(\"\\n=== MAPE Comparison ===\\n\")\nprint(mape_table)\n\n# =============================================================================\n# SAFETY STOCK: SS = z * RMSE(forecast_error) * sqrt(L)\n# z = 1.645 (95% SL), L = 1 month\n# =============================================================================\n\nz_val <- 1.645\nlead_time <- 1\n\nsafety_stock <- comparison |>\n  group_by(sku) |>\n  summarise(\n    avg_demand = round(mean(demand)),\n    sales_rmse = sqrt(mean((demand - sales_forecast)^2, na.rm = TRUE)),\n    stat_rmse  = sqrt(mean((demand - stat_forecast)^2, na.rm = TRUE)),\n    .groups = \"drop\"\n  ) |>\n  mutate(\n    sales_ss         = round(z_val * sales_rmse * sqrt(lead_time)),\n    stat_ss          = round(z_val * stat_rmse * sqrt(lead_time)),\n    ss_reduction_pct = round((1 - stat_ss \/ sales_ss) * 100, 1),\n    units_saved      = sales_ss - stat_ss\n  )\n\ncat(\"\\n=== Safety Stock Impact ===\\n\")\nprint(safety_stock)\n\nunit_cost <- 25\ntotal_sales_ss <- sum(safety_stock$sales_ss)\ntotal_stat_ss  <- sum(safety_stock$stat_ss)\ncat(paste0(\"\\nPortfolio: \", total_sales_ss, \" -> \", total_stat_ss,\n           \" units (\", round((1 - total_stat_ss \/ total_sales_ss) * 100, 1),\n           \"% reduction,<\/code><\/pre>\n
      <\/pre>\n
      quot;, (total_sales_ss - total_stat_ss) * unit_cost, \" saved at $25\/unit)\\n\")) # ============================================================================= # CHARTS (6 images) \u2014 see generate_sop_images.R for full plotting code # ============================================================================= # Image 1: sop_demand_patterns.png (800x700) \u2014 6 faceted demand panels # Image 2: sop_showdown_seasonal.png (800x500) \u2014 SKU-A actual vs forecasts # Image 3: sop_showdown_trending.png (800x500) \u2014 SKU-B actual vs forecasts # Image 4: sop_showdown_portfolio.png (800x500) \u2014 Grouped bar MAPE comparison # Image 5: sop_safety_stock_impact.png(800x500) \u2014 Safety stock paired bars # Image 6: sop_toolkit_tiers.png (800x400) \u2014 Progressive roadmap # Full plotting code in generate_sop_images.R (lines 250-504) <\/code><\/pre>\n<\/details>\n
      References<\/h2>\n
        \n
      1. Gartner (2024). S&OP Maturity Model: How Supply Chain Leaders Drive Cross-Functional Alignment<\/em>. Gartner Research.<\/li>\n
      2. Hyndman, R.J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice<\/em>, 3rd edition. OTexts. https:\/\/otexts.com\/fpp3\/<\/a><\/li>\n
      3. Silver, E.A., Pyke, D.F., & Thomas, D.J. (2017). Inventory and Production Management in Supply Chains<\/em>, 4th edition. CRC Press.<\/li>\n
      4. Syntetos, A.A., & Boylan, J.E. (2005). „The accuracy of intermittent demand estimates.“ International Journal of Forecasting<\/em>, 21(2), 303-314.<\/li>\n
      5. Lapide, L. (2005). „Sales and operations planning Part III: A diagnostic model.“ Journal of Business Forecasting<\/em>, 24(1), 13-16.<\/li>\n
      6. Moon, M.A. (2018). Demand and Supply Integration: The Key to World-Class Demand Forecasting<\/em>, 2nd edition. De Gruyter.<\/li>\n
      7. Gilliland, M. (2010). The Business Forecasting Deal: Exposing Myths, Eliminating Bad Practices, Providing Practical Solutions<\/em>. Wiley.<\/li>\n
      8. Thom\u00e9, A.M.T., Scavarda, L.F., Fernandez, N.S., & Scavarda, A.J. (2012). „Sales and operations planning: A research synthesis.“ International Journal of Production Economics<\/em>, 138(1), 1-13.<\/li>\n
      9. Tuomikangas, N., & Kaipia, R. (2014). „A coordination framework for sales and operations planning (S&OP): Synthesis from the literature.“ International Journal of Production Economics<\/em>, 154, 243-262.<\/li>\n
      10. Grimson, J.A., & Pyke, D.F. (2007). „Sales and operations planning: An exploratory study and framework.“ The International Journal of Logistics Management<\/em>, 18(3), 322-346.<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"
        Only 15% of companies run S&OP successfully. For the other 85%, here’s a data-driven toolkit that lets production planners bypass the broken process.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13,115],"tags":[126,122,205,125,128,203,119,201,202,204],"class_list":["post-1297","post","type-post","status-publish","format-standard","hentry","category-data-science","category-supply-chain-management","tag-arima","tag-demand-forecasting","tag-demand-planning","tag-ets","tag-forecast-accuracy","tag-mape","tag-r-programming","tag-sop","tag-safety-stock","tag-statistical-forecasting"],"_links":{"self":[{"href":"https:\/\/inphronesys.com\/index.php?rest_route=\/wp\/v2\/posts\/1297","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/inphronesys.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/inphronesys.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/inphronesys.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/inphronesys.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1297"}],"version-history":[{"count":1,"href":"https:\/\/inphronesys.com\/index.php?rest_route=\/wp\/v2\/posts\/1297\/revisions"}],"predecessor-version":[{"id":1298,"href":"https:\/\/inphronesys.com\/index.php?rest_route=\/wp\/v2\/posts\/1297\/revisions\/1298"}],"wp:attachment":[{"href":"https:\/\/inphronesys.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1297"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/inphronesys.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1297"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/inphronesys.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1297"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}