Best Practices for Sales Forecasting

Google just open-sourced a time series foundation model. It works with any data without training. Traditional forecasting models need to be trained on your specific dataset before they can predict anything. Google's TimesFM works differently. you give it historical data, and it generates forecasts out of the box. This is possible because they trained the model on 100 billion real-world time-points across domains like traffic, weather, and demand forecasting. Key features: - supports up to 16K context length for deeper historical coverage - built-in probabilistic forecasting with quantile predictions - works with both PyTorch and JAX It currently sits at the top of GIFT-Eval, the standard benchmark for time-series forecasting. If you work with demand forecasting, financial data, or any time-series problem, this is worth exploring. I've shared the link to the GitHub repo in the first comment. ____ Share this with your network if you found this insightful ♻️ Follow me (Akshay Pachaar) for more insights and tutorials on AI and Machine Learning!

(FMCG Blueprint) Sales forecasting in FMCG is both an art and a science. Let’s break it down using some basic matrices with a relatable example. Imagine we’re working for a brand that sells a spicy instant noodle, “HotBowl Ramen”. 1. Historical Sales Data (Your Crystal Ball) The first step is to look at past sales. For example: Month Sales (Units) January 10,000 February 11,000 March 10,500 April 12,000 Now, let’s assume you notice a 5% growth trend every month. For May, you might forecast: May Sales = April Sales * (1 + Growth Rate) = 12000 * (1 + 0.05) = 12600 Tip: This works well unless your sales suddenly nosedive because people discovered a new health fad: “No-Spice Life!” 2. Seasonality (Your FMCG Calendar) People eat more noodles in winter because “cozy food” vibes. Let’s adjust for seasonality: • Winter months: Add 10% • Summer months: Subtract 15% If your May forecast is 12,600 units but May is peak summer, adjust like this: Adjusted Sales = Base Sales * (1 - 0.15) = 12600*0.85 = 10,710 Reality Check: Your product is spicy. Some brave souls will still eat it even in May, sweating like they’re in a sauna. 3. Market Dynamics (Your Frenemy) Suppose your competitor, “MildBowl Ramen,” launches a huge promotion in May. You estimate a 10% impact on your sales. Final Sales Forecast = Adjusted Sales * (1 - 0.1) = 10710*0.9 = 9,639 4. Promotional Impact (Buy One, Cry One Free?) Now, your marketing team swoops in with a “Buy 1 Get 1 Free” promo. Promotions can boost sales by 20%, so: Promo Adjusted Sale = 9639*1.2 =11,566.8 Realistic Case Summary Step Forecasted Sales Base Sales Forecast 12,600 Seasonality Adjustment 10,710 Competitor Impact 9,639 Promo Impact 11,566 Funny Perspective Imagine your boss: • Before Forecast: “We need 15,000 units this month!” • After Your Analysis: “Hmm… okay, but let’s add another promo to reach 12,000 at least!” Your real hero? The customer who eats your spicy noodles even in May, sweating but happy. Moral: Forecasting is like cooking ramen—balance your ingredients (data) and adjust for taste (market trends)!

Modeling something like time series goes past just throwing features in a model. In the world of time series data, each observation is associated with a specific time point, and part of our goal is to harness the power of temporal dependencies. Enter autoregression and lagging -  concepts that taps into the correlation between current and past observations to make forecasts.  At its core, autoregression involves modeling a time series as a function of its previous values. The current value relies on its historical counterparts. To dive a bit deeper, we use lagged values as features to predict the next data point. For instance, in a simple autoregressive model of order 1 (AR(1)), we predict the current value based on the previous value multiplied by a coefficient. The coefficient determines the impact of the past value on the present one only one time period previous. One popular approach that can be used in conjunction with autoregression is the ARIMA (AutoRegressive Integrated Moving Average) model. ARIMA is a powerful time series forecasting method that incorporates autoregression, differencing, and moving average components. It's particularly effective for data with trends and seasonality. ARIMA can be fine-tuned with parameters like the order of autoregression, differencing, and moving average to achieve accurate predictions. When I was building ARIMAs for econometric time series forecasting, in addition to autoregression where you're lagging the whole model, I was also taught to lag the individual economic variables. If I was building a model for energy consumption of residential homes, the number of housing permits each month would be a relevant variable. Although, if there’s a ton of housing permits given in January, you won’t see the actual effect of that until later when the houses are built and people are actually consuming energy! That variable needed to be lagged by several months. Another innovative strategy to enhance time series forecasting is the use of neural networks, particularly Recurrent Neural Networks (RNNs) or Long Short-Term Memory (LSTM) networks. RNNs and LSTMs are designed to handle sequential data like time series. They can learn complex patterns and long-term dependencies within the data, making them powerful tools for autoregressive forecasting. Neural networks are fed with past time steps as inputs to predict future values effectively. In addition to autoregression in neural networks, I also used lagging there too! When I built an hourly model to forecast electric energy consumption, I actually built 24 individual models, one for each hour, and each hour lagged on the previous one. The energy consumption and weather of the previous hour was very important in predicting what would happen in the next forecasting period. (this model was actually used for determining where they should shift electricity during peak load times). Happy forecasting!

CEO: Our margins are getting tighter. FP&A: Let’s cut costs. CEO: We’re missing revenue targets. FP&A: Let’s reforecast. CEO: Our cash flow is unpredictable. FP&A: Let’s track it closer. CEO: We’re losing market share. FP&A: Let’s adjust assumptions. This is how finance becomes a back-office function. And it’s why most FP&A teams get ignored in strategy meetings. Instead, try this: 1. Turn data into decisions, not just reports CEOs don’t need more charts. They need answers. If your reports don’t drive action, they’re just noise. FP&A teams that translate numbers into clear next steps get a seat at the table. 2. Make forecasting dynamic, not static Annual budgets are already outdated by Q2. Winning teams run rolling forecasts that adapt in real-time, using leading indicators to predict what’s next, before the business feels the impact. 3. Use capital as a competitive advantage The best companies don’t just cut costs, they allocate capital better. Instead of reacting to margin pressure with blanket cuts, double down on high-ROI opportunities and phase out low-value spending. 4. Speak the language of business Finance gets ignored when it talks in numbers, not outcomes. Saying, “Gross margin fell by 2%” misses the mark. Saying, “Optimizing pricing can recover $5M in profit next quarter” gets action. 5. Don’t wait for leadership to ask The best FP&A teams don’t wait. They anticipate challenges, model different scenarios, and push strategic moves before the company is forced to react. Influence happens when finance drives the conversation, not follows it. The FP&A teams winning in 2025 aren’t managing costs. They’re out-executing their competitors. FP&A sees what’s coming first. Follow Erik Lidman for FP&A insights.

Top-Down vs. Bottom-Up Forecasting 📊 Which approach should you choose for your next forecast? I see companies get this wrong all the time...and it costs them big. ➡️ TOP-DOWN FORECASTING This one starts with the big picture. Market size, growth targets, high-level assumptions about where you fit in the world. "We want 5% market share...equals $10M in revenue." Fast and directional? Absolutely. Executives and VCs eat this up during early-stage planning because it ties directly to strategy. But there's a catch... ✅ Quick to build ✅ Aligns with big-picture strategy ❌ Can be overly optimistic ❌ Misses execution details ➡️ BOTTOM-UP FORECASTING Now this is where things get real. You're building from actual internal data...team capacity, sales pipeline, product usage. "Each rep closes 5 deals per month...20 reps = $1.2M monthly." It's grounded in what you actually have, not what you hope to achieve. No wishful thinking allowed. ✅ Realistic and execution-focused ✅ Helps with hiring, spend, and capacity plans ❌ Slower to build ❌ Can miss strategic targets if not guided top-down ➡️ SO WHICH ONE SHOULD YOU USE? Here's my take... Early-stage companies start top-down for fundraising and strategic planning. Makes sense. But once you have real operational data? Bottom-up becomes way more accurate for running the business day-to-day. The best forecasts combine both. Start top-down to set ambitious targets, then validate with bottom-up to make sure your plan is actually achievable. === Most finance teams pick one and stick with it...but that's a mistake. What forecasting approach has worked best for you? Let me know in the comments below 👇

"Just curious, how's your forecast looking?" My CEO friend asked me. The weekly forecast review. The monthly pipeline call. The quarterly business review. All centered around one flawed model: Asking reps to predict the future based on gut feeling. "50% chance of closing." "Strong verbal commitment." "Just waiting on final approval." These phrases hide a painful truth: We have no idea what's actually happening inside our deals. I changed how we forecast last quarter: Instead of: "How do you FEEL about this deal?" We now ask: "What have they actually DONE?" - Has the economic buyer viewed pricing? - Have technical stakeholders reviewed security docs? - Have end users looked at implementation plans? - Is the champion actively sharing content internally? Behavior doesn't lie. Words do. We tracked content engagement across 200+ deals: Closed deals: Prospects engaged 7+ times in final two weeks Lost deals: Engagement dropped to 0-1 interactions before going dark The deals your team is most confident about? Often the ones with the least actual buyer engagement. Here's how we transformed our approach: Every opportunity now has a digital space where we can see: - Exactly who is engaging with what content - Which stakeholders are involved (even ones we haven't met) - Where deals are getting stuck - When interest spikes or drops Our forecast accuracy improved INSANELY. Stop asking reps what they "think" will happen. Start measuring what buyers are actually doing. The best indication of deal health isn't what prospects tell you. It's how they behave when you're not watching. Do you know what your buyers are really doing? Or are you still forecasting based on feelings? Agree?

Sales forecasting isn’t just about projecting revenue. It’s about understanding what drives revenue. Here are a few examples. (1) Price x Volume I usually don't separate sales into rates and units because of the extensive detail required. Most of my forecasts are all-in sales of price x volume, or rates x units. It's usually 'good enough' and balances accuracy with effort. But you know it's not always appropriate. If you want precision, or scenario modeling, you'll likely need to break these down further. If prices aren't fixed or demand is dynamic, you'll likely need to deliver a more detailed forecast. (2) Include/Exclude Toggles Sales pipelines often contain CRMs with customers at different stages in the sales cycle. Including them, or applying % volume reductions based upon uncertainty, can distort the sales forecast. In my models, I like to include toggles (similar to the checkboxes you see here) that allow for the inclusion/exclusion of sales depending on (a) scenarios, or (b) the stage of the sales process. This lets you easily change your sales forecast without corrupting your formulas. (3) Top-Down Forecasts Not all forecasts can (or should) be bottoms-up. In this example, the company has a huge opportunity with “NFL Confidential” customer. This customer may or may not be landed, which is why there's an include/exclude toggle. FP&A also included macro-level assumptions for the events that will drive sales up or down. It's a top-down estimate, modeled from known business events (the NFL playoffs) from Q4 to Q1. Sales ramp up slightly, then significantly, before they come back down. (4) Customer Concentration This company may be eager to land an NFL team as a customer, as it's both a strategic and financial play. On the strategic side, the company can get greater market exposure. On the financial side, it brings $5.3 million to the top line. But this amounts to 26.5% of total sales, huge concentration. So there are questions to ask: Can the company effectively manage this higher volume? How does this new focus disrupt other operations? Will new roles need to be filled to accommodate the customer? Are different machines and new capex necessary to service the customer? Does the company have the liquidity to obtain raw materials? What timing for deposits and billings allows the company to cash flow? Remember: sales forecasting isn’t just about projecting revenue. It's about understanding the drivers and implications. When sales forecasting becomes a joint effort between sales and FP&A, you get a far more thoughtful planning process.

Hey Salespeople: While I feel strongly that probabilities for sales stages should be set based on historical conversion rates, I discovered a different option today about probabilities for forecast categories when speaking with Rachel Chan and Alexandre Perrin-Delort from GitGuardian. Instead of using historical rates, assign probabilities and ask, 'What must be true to achieve this probability?' For instance... Commit (90%) --> Verbal agreement from the EB --> Commercial terms are locked --> Legal/procurement is in motion (or stricter, order form is out of signature)  Most LIkely (60%) --> Champion has gotten directional approval from the EB for solution at the proposed budget --> Procurement and legal have been identified and engaged --> PoC/PoV is complete and champion has selected you as vendor of choice Best Case (30%) --> You have identified and met with the EB (preferably via intro from the Champion) --> Prospect has engaged with a proposal or business case (preferably co-developed with Champion) --> Decision timeline by close date is plausible Pipeline (10%) --> Opp is qualified with confirmed pain --> Next meeting is scheduled Omit (0%) Either (a) discovery meeting scheduled (b) discovery meeting held but no next meeting scheduled High quality forecasting requires not only strict definitions but also rigorous & frequent inspection. One should test these historical conversion rates against the assigned probabilities and tighten the criteria as needed. All of this prompts another questions: Should one use stages, forecast categories, or both? The criteria above are related to stage entry/exit criteria. For SMB, stages alone are sufficient since it is not a good use of manager time to review every deal. In MM, you'll want both. In ENT, forecast categories alone are sufficient since teams will scrutinize every deal. More generally, I recommend using both stages & forecast categories in MM and ENT. Stages give you a forecast based on historical conversion rates. Forecast categories give you a forecast based on informed judgement. One can then build a forecast a mixture of the two. For the record, I'm not a fan of periodically gathering "call" numbers from reps and managers. Yes, this does have some psychological commitment impact but not enough to justify the time & effort.

Your pipeline is lying to you And it's costing you millions in missed forecasts. The problem isn't your CRM. It's how you're defining "opportunity." Most sales teams call something an opportunity the moment a prospect shows interest. But interest isn't intent. After tracking thousands of deals, I've found that 68% of "opportunities" in most pipelines will never result in a purchase decision. They're just conversations disguised as deals. Here's how to clean up your pipeline and start forecasting accurately: An opportunity only exists when these three elements are present: 1️⃣ Documented business impact  Not "they have problems" but "solving this will save them $X or generate $Y" 2️⃣ Defined evaluation process  Not "they're interested" but "here's their timeline, stakeholders, and decision criteria" 3️⃣ Accessible buying team  Not "I found a champion" but "I can reach the people who actually influence this decision" Without all three, you have a conversation, not an opportunity. The companies with the most accurate forecasts are ruthless about this distinction. They'd rather have a smaller, cleaner pipeline than a bloated one full of wishful thinking. This forces reps to qualify harder upfront and focus their energy on deals that can actually close.

You can't treat every forecast the same. More uncertainty means more risk, and you want to deal with it correctly. After building forecasting models at P&G, Unilever, and Squarespace, I've learned there are three ways to manage uncertainty: 𝟭) 𝗔𝘃𝗼𝗶𝗱 𝗔𝘀𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻 𝗦𝘁𝗮𝗰𝗸𝗶𝗻𝗴 The more uncertainty, the fewer assumptions you should include. Why? Because if you add multiple variables on top of each other, their margin of error multiplies. If you base the forecast on many assumptions, it's nearly impossible to determine which one was accurate and which wasn't. So, keep your models as simple as possible. Isolate the variables. You can always add additional assumptions later once you better understand the correlations. 𝟮) 𝗥𝘂𝗻 𝗪𝗵𝗮𝘁-𝗜𝗳 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀 It's your job as a finance leader to quantify the risk of a forecast. The easiest way to do that is by changing individual inputs and noting how much impact that has on the forecast. For example, if a 5% price change affects the revenue forecast by 25%, that's a major risk you'll need to call out. 𝟯) 𝗦𝗵𝗼𝘄 𝗮 𝗥𝗮𝗻𝗴𝗲 Sometimes analysts make the mistake of assuming ranges make it look like they aren't confident in their forecast. But a well-measured range is critical for two reasons: One, it shows the order of magnitude of risk. Your CFO knows what's a conservative estimate to communicate to investors. Two, it enables scenario planning. Leaders can plan contingency measures if results are at the lower end of the range. 𝗜𝗻 𝘀𝘂𝗺, 𝘁𝗼 𝗺𝗮𝗻𝗮𝗴𝗲 𝘂𝗻𝗰𝗲𝗿𝘁𝗮𝗶𝗻𝘁𝘆 𝗶𝗻 𝗮 𝗺𝗼𝗱𝗲𝗹: 1. Reduce the number of assumptions 2. Estimate the risk by running sensitivity analysis 3. Provide ranges instead of point estimates Which approach do you find most useful? Comment below 👇 -Christian Wattig 📌 Get my 𝗙𝗶𝗻𝗮𝗻𝗰𝗶𝗮𝗹 𝗠𝗼𝗱𝗲𝗹𝗶𝗻𝗴 𝘁𝗲𝗺𝗽𝗹𝗮𝘁𝗲 + 𝟰𝟲 𝗯𝗲𝘀𝘁 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲𝘀 (free) here: https://lnkd.in/eBAmSF_6