Forecasting & Modelling prompts

Help me build a forecast.

WHAT I AM FORECASTING: [THE METRIC]
HISTORICAL DATA:
"""
[PASTE - PERIODS AND VALUES]
"""
HOW FAR FORWARD: [PERIODS]
WHAT THE FORECAST IS FOR: [THE DECISION]
WHAT I KNOW ABOUT WHAT DRIVES THIS: [CAUSAL FACTORS]
WHAT IS CHANGING: [ANYTHING PLANNED OR EXPECTED]

Produce:

1. WHAT THE HISTORY SHOWS - the trend, the variation around it, any seasonality, and any structural breaks where the pattern changed. Identify breaks explicitly; forecasting through a break using pre-break data is the most common forecasting error.

2. HOW MUCH HISTORY IS USABLE - if something changed partway through, only the post-change data is informative. Say how many periods you are actually forecasting from, which is often far fewer than the total.

3. THE BASELINE FORECAST - the simplest defensible method for this data: last value carried forward, average, trend extrapolation, or seasonal naive. Start here, because a simple baseline is frequently as accurate as anything elaborate and always easier to explain.

4. THE RANGE, NOT THE NUMBER - the forecast as low, likely and high, based on the historical variation. A single-number forecast is a false statement about the future. Say how wide the range should be given how variable the history is.

5. THE ASSUMPTIONS - each explicitly: that the trend continues, that seasonality repeats, that nothing changes structurally, and that the drivers I named continue to behave as they have. List them, because these are what the forecast actually rests on.

6. WHAT WOULD BREAK IT - the specific events that would make this forecast wrong, and roughly how wrong.

7. THE PLANNED CHANGES - I told you what is changing. The historical data cannot inform their effect. Handle them as an explicit adjustment with its own stated assumption, separate from the statistical forecast, rather than folding them in invisibly.

8. THE HORIZON WARNING - forecast accuracy degrades with distance. Say at what horizon this forecast stops being informative, given the variability in my data.

9. HOW TO USE IT - given my decision: does the decision change across the range? If it is the same at the low and high case, stop forecasting and decide.

10. HOW TO CHECK IT - record the forecast now and compare later. Forecasts that are never scored never improve.

Do not produce a precise-looking number. Produce a range and its assumptions.

Help me build scenarios.

THE SITUATION: [WHAT YOU ARE PLANNING FOR]
THE DECISION: [WHAT YOU NEED TO DECIDE]
TIME HORIZON: [HOW FAR OUT]
WHAT IS UNCERTAIN: [THE KEY UNKNOWNS]
WHAT YOU KNOW: [THE FIXED FACTS]
WHAT YOU CAN CONTROL: [YOUR LEVERS]

Produce:

1. THE CRITICAL UNCERTAINTIES - of everything uncertain, the two that most affect the outcome and are least predictable. Rank all my stated unknowns by (impact x unpredictability) and pick the top two. Things that are uncertain but low-impact do not need scenarios.

2. THE SCENARIO SET - built from combinations of those two uncertainties. Three or four scenarios, each with a short memorable name. For each:
   - What is true in this world, described concretely
   - How you would recognise it early
   - What it means for my decision
   - What would be the right move in this scenario alone

   Scenarios must be genuinely different, plausible, and internally consistent. Avoid the common failure of good case, bad case and middle case - that is a range, not a set of scenarios, and it produces no insight.

3. THE ROBUST CHOICES - what is the right thing to do across all scenarios. This is the most valuable output: decisions that work regardless of which future arrives.

4. THE SCENARIO-DEPENDENT CHOICES - what you should not commit to yet because the right answer differs by scenario. And what would let you defer them cheaply.

5. THE SIGNPOSTS - early indicators that would tell you which scenario is developing, and roughly when you would see each. This is what turns scenarios from an exercise into a monitoring system.

6. THE OPTIONS TO PRESERVE - cheap things to do now that keep possibilities open: a small pilot, a shorter contract, a reversible commitment. Optionality is worth paying for when uncertainty is high.

7. THE UNACCEPTABLE OUTCOME - the scenario you could not survive, however unlikely. What would protect against it, and what that protection costs.

8. WHAT I AM ASSUMING IS FIXED that might not be. The things I listed as known are themselves assumptions; challenge the shakiest.

Do not produce a forecast. The point is to prepare for several futures, not to pick one.

Help me estimate this.

WHAT I NEED TO ESTIMATE: [THE QUANTITY]
WHY: [THE DECISION IT INFORMS]
WHAT I KNOW: [ANY RELEVANT FACTS OR FIGURES]
HOW ACCURATE IT NEEDS TO BE: [order of magnitude / within 50% / better]

Produce:

1. THE DECOMPOSITION - break the quantity into components you can estimate more confidently. Keep breaking down until each piece is something you can reason about or look up. This is the whole technique: nobody can estimate the target directly, and everyone can estimate its parts.

2. THE CALCULATION CHAIN - each component with: the estimate, the basis (a known fact, a reasonable assumption, or a guess), and its uncertainty. Show the arithmetic.

3. TWO INDEPENDENT ROUTES - estimate the same quantity a second way, using different components. If the two agree within your required accuracy, confidence increases substantially. If they diverge wildly, at least one chain contains an error worth finding. This step is what separates an estimate from a guess.

4. THE RANGE - low, likely and high. Build it from the uncertainty in each component rather than by putting a margin around the point estimate. Note that uncertainties do not simply add; a chain of four components each uncertain by a factor of two does not produce a factor-of-sixteen range in practice.

5. THE SENSITIVITY - which component most drives the answer. Usually one or two dominate, and those are the only ones worth spending effort to improve.

6. THE SANITY CHECKS - constraints the answer must satisfy. Is it consistent with any total it must fit inside? Is it plausible per person, per day, or per unit? Does it imply anything absurd? An estimate that implies more hours than exist in a day has an error in it.

7. THE ANCHOR CHECK - if I gave you a figure I already had in mind, assess whether the decomposition supports it or whether I have reasoned backwards from it.

8. WHAT WOULD MOST IMPROVE THIS - the one fact worth looking up, given the sensitivity analysis.

9. IS IT GOOD ENOUGH - given my stated accuracy requirement and decision, does this estimate support the decision? If the decision is the same across the whole range, stop.

Be explicit about which numbers are known, assumed, or guessed. A chain of guesses presented as a calculation is worse than saying you do not know.

Review this forecast.

THE FORECAST:
"""
[PASTE - THE NUMBERS, THE PERIODS, AND ANY STATED METHOD]
"""

WHAT IT IS FORECASTING: [THE METRIC]
WHO PRODUCED IT: [AND THEIR INTEREST IN THE ANSWER]
WHAT DECISION DEPENDS ON IT: [THE STAKES]
HISTORICAL DATA IF AVAILABLE: [PASTE]

Produce:

1. THE ASSUMPTIONS - stated and unstated. Every forecast rests on assumptions; a forecast that does not state them is not reviewable. List what must be true for these numbers to hold, and mark which were stated and which you had to infer.

2. THE GROWTH CHECK - what growth rate does this imply, period on period? Compare it to the historical rate if I gave you one, and to what is plausible for this kind of thing. Forecasts that imply sustained growth far above anything historically achieved need justification that is usually absent.

3. THE HOCKEY STICK TEST - does the forecast show flat or declining recent history followed by sharp improvement? If so, what specifically causes the inflection? An unexplained inflection is the single most common sign of a forecast built backwards from a desired conclusion.

4. THE CONSTRAINT CHECK - does the forecast respect real limits? Market size, capacity, headcount required, physical or regulatory constraints. Compute what the forecast implies in units other than the headline figure: customers needed, units per day, staff required. Implied figures are where over-optimism becomes visible.

5. THE MISSING COSTS - if it is a financial forecast, what scales with growth that is not shown: support, infrastructure, hiring, churn at higher volumes.

6. THE UNCERTAINTY - is there any? A single-line forecast with no range is a statement of intent, not a projection. If no range is given, ask for one.

7. THE INCENTIVE CHECK - I told you who made it. Do they benefit from an optimistic number? This does not make it wrong, but it determines how much independent verification you should want.

8. THE BASE RATE - how do forecasts of this type usually perform? Say honestly that most such forecasts overstate, and by roughly how much, if that is the pattern.

9. THE QUESTIONS TO ASK - specific and answerable, phrased neutrally. The most useful is usually: what would have to be true for this, and what happens if it is not?

10. THE VERDICT - credible / optimistic but defensible / not supported by the stated basis.

Help me model the financial effect of this decision.

THE DECISION: [WHAT YOU ARE CONSIDERING]
WHAT IT COSTS: [UPFRONT AND ONGOING]
WHAT IT SHOULD RETURN: [THE BENEFIT YOU EXPECT]
CURRENT SITUATION: [THE BASELINE - what happens if you do nothing]
TIME HORIZON: [PERIOD]
WHAT YOU KNOW AND WHAT YOU ARE GUESSING: [BE EXPLICIT]

Produce:

1. THE BASELINE - what happens if you do nothing. Every comparison is against this, and it is frequently omitted, which makes the proposed option look better than it is. Model the do-nothing case first.

2. THE COST MODEL - all of it, over the horizon:
   - Upfront and ongoing
   - Staff time, costed
   - The things people forget: training, migration, integration, support, maintenance, the cost of running both old and new during transition, and the cost of stopping if it does not work
   - Opportunity cost: what does not happen because resources went here

3. THE BENEFIT MODEL - with the mechanism stated. Not 'improves efficiency' but the specific chain: this changes, which means that, which is worth this much. Where a link in the chain is an assumption, mark it.

4. THE TIMING - when costs fall and when benefits arrive. Benefits almost always arrive later and more slowly than expected. Model a realistic ramp rather than full benefit from day one.

5. THE NET PICTURE - period by period, cumulative, and the point at which it turns positive if it does.

6. THE SENSITIVITY - the two or three assumptions that most affect the result. Show the outcome at pessimistic, expected and optimistic values for each. This matters more than the central estimate.

7. THE BREAK-EVEN - what would each key assumption need to be for this to merely break even? Stating it this way is often more informative than the projection: 'this needs a 12% improvement to break even, and we currently see 4%'.

8. WHAT IS NOT IN THE MODEL - effects that are real but not quantified: risk reduction, morale, flexibility, strategic position. Name them so they are considered rather than ignored, and do not invent numbers for them.

9. THE HONEST SUMMARY - does this stack up? If it only works under optimistic assumptions on every variable simultaneously, say so.

I am not asking for financial advice, only for this arithmetic laid out with its assumptions visible.

Help me assess my forecasting.

MY PAST PREDICTIONS:
"""
[PASTE - WHAT YOU PREDICTED, WHEN, WITH WHAT CONFIDENCE, AND WHAT ACTUALLY HAPPENED]
"""

DOMAIN: [WHAT THESE ARE ABOUT]
HOW I MADE THEM: [YOUR METHOD, if any]

Produce:

1. THE SCORECARD - each prediction: what you said, your stated confidence, the outcome, and whether it was right. Where a prediction was too vague to score, say so - that is itself the most important finding, because unfalsifiable predictions feel like forecasting and are not.

2. THE CALIBRATION CHECK - group predictions by stated confidence. Of the ones you were 70% confident about, roughly 70% should have happened. More than that means you are underconfident; fewer means overconfident. Overconfidence is the normal result.

3. THE DIRECTIONAL BIAS - do you systematically predict too optimistically, too pessimistically, too much change, or too much continuity? Most people over-predict change in the short term and under-predict it over longer horizons.

4. WHAT YOU GOT RIGHT AND WHY - and importantly, whether you were right for the reason you gave. Being right for the wrong reason is not skill and will not repeat.

5. WHAT YOU GOT WRONG - grouped by cause:
   - Missing information you could have had
   - Information you had and discounted
   - A reasoning error
   - Genuine unpredictability
   Only the first three are improvable. Separating them stops you learning the wrong lesson from bad luck.

6. THE HINDSIGHT WARNING - for the ones you got wrong, resist the reconstruction that it was obvious. It was not, or you would have predicted it. Assess whether the information available at the time actually supported the correct answer.

7. THE VAGUENESS PROBLEM - predictions phrased so they cannot be wrong. Quote them and rewrite each as something scoreable, with a date and a threshold.

8. THE PATTERN - the one systematic error that recurs.

9. HOW TO FORECAST BETTER in this domain: record predictions with dates, confidence levels and reasoning; make them specific enough to score; and check them on a schedule.

Be direct about overconfidence. It is the normal state and it is only correctable once measured.

Help me plan capacity.

WHAT I NEED TO SIZE: [STAFF / STOCK / SERVERS / SPACE / APPOINTMENTS]
DEMAND HISTORY:
"""
[PASTE - PERIODS AND VOLUMES]
"""
EXPECTED CHANGE: [WHAT YOU THINK WILL HAPPEN]
PERIOD TO PLAN FOR: [TIMEFRAME]
COST OF TOO LITTLE: [WHAT HAPPENS IF YOU ARE SHORT]
COST OF TOO MUCH: [WHAT SPARE CAPACITY COSTS]
LEAD TIME TO ADD CAPACITY: [HOW LONG IT TAKES]

Produce:

1. THE DEMAND PATTERN - average, and more importantly the variation: peaks, troughs, seasonality, and day-of-week or time-of-day effects. Capacity planned to the average fails at the peak, and this is the single most common capacity mistake.

2. THE PEAK ANALYSIS - how high do peaks go, how often, and how long do they last? Plan against a stated percentile of demand rather than the mean, and say which percentile and why.

3. THE ASYMMETRY - I told you the cost of too little and too much. These are almost never equal. Where being short is much more expensive than being over-provisioned, plan for a higher percentile, and vice versa. State explicitly which way this should lean and by how much.

4. THE UTILISATION TRAP - capacity planned for high average utilisation produces long waits, because arrivals are irregular. A system running at 90% average utilisation has queues far longer than one at 70%, not slightly longer. If waiting matters here, say what utilisation target is appropriate.

5. THE LEAD TIME PROBLEM - you said how long capacity takes to add. That means you must decide based on demand that far ahead, not current demand. State the decision horizon and what signal should trigger adding capacity.

6. THE RECOMMENDATION - the capacity level, with reasoning, and the range it comfortably serves.

7. THE FLEXIBILITY OPTIONS - ways to handle peaks without permanent capacity: overflow arrangements, temporary staff, queueing, demand shifting, or an agreement with a partner. Flexible capacity is often cheaper than peak capacity.

8. THE TRIGGERS - the specific observable conditions that mean add capacity now, and the ones that mean reduce it. Decided in advance, because in the moment you will be reacting to the last week's noise.

9. WHAT COULD GO WRONG - a demand change larger than planned, capacity becoming unavailable, or a step change that the history does not predict.

Help me build a 13-week cash flow forecast.

BUSINESS: [DESCRIBE]
OPENING CASH BALANCE: [AMOUNT]
REGULAR RECEIPTS: [CUSTOMERS, AMOUNTS, TIMING, PAYMENT TERMS]
REGULAR PAYMENTS: [PAYROLL, RENT, SUPPLIERS, LOANS, TAXES - AMOUNTS AND DATES]
ONE-OFF ITEMS: [LIST]
TOOL: [Excel / Google Sheets]

Produce:
1. THE LAYOUT - rows (receipts by type, payments by type, net flow, opening and closing balance) and 13 weekly columns
2. FORMULAS for my tool, with cell references explained
3. ASSUMPTIONS TAB - every assumption listed with where it is used
4. A WARNING LINE - conditional formatting when closing cash falls below [MINIMUM BALANCE]
5. HOW TO UPDATE IT WEEKLY - replace forecast with actuals and roll forward
6. The 3 levers if a shortfall appears (accelerate receipts, delay payments, short-term funding) with what to check for each

Rules: do not invent amounts; use [AMOUNT] placeholders for anything I did not give.