Forecasting Under Uncertainty: Where Machine Prediction Beats Judgement, and Where It Does Not

of clean monthly history is a practical minimum for a seasonal sales model

3 years

is the horizon of the cash forecast most Nepali companies need first

13 weeks

scenarios with triggers are worth more than one precise number

2 to 3

Every business in Nepal forecasts. The distributor deciding how much to order before Dashain is forecasting. The hotelier setting rates for the spring season is forecasting. The finance head who tells the bank that the overdraft will be regularised by Falgun is forecasting. Most of it is done in the owner's head, informed by years of experience, and much of it is good.

The claim made for artificial intelligence is that a machine can do this better. Sometimes it can, decisively. Sometimes it cannot, and a company that trusts the machine at the wrong moment will be worse off than one that never bought it. The skill lies in knowing which situation you are in.

This is the eighth article in our series on artificial intelligence for business in Nepal, and the third on finance, after the CFO's office in 2030 and redesigning the close.

What the evidence says

Research on forecasting is unusually mature, and three findings from decades of empirical work, including the long-running M forecasting competitions, are worth knowing.

Simple statistical methods are hard to beat. For many business series, a method as plain as "the same month last year, adjusted for recent trend" performs close to far more elaborate models. Complexity earns its keep only with large volumes of data and many related series.

Combinations beat individual forecasts. Averaging several methods, or a model with informed judgement, is more accurate on average than any single one. This is among the most robust findings in the field.

People adjust too often and too optimistically. Studies of judgemental adjustments to statistical forecasts find that large adjustments made for a specific, known reason tend to improve accuracy, while frequent small adjustments tend to make it worse, and that adjustments lean upward.

Modern machine learning has added real capability at scale: retailers forecasting millions of item-store combinations, lenders estimating default, utilities forecasting load. The most recent M competitions were won by machine learning methods on exactly that kind of large, hierarchical data. For a company with forty products and five years of monthly history, the older lessons still hold.

Where do the large consulting firms stand? Their surveys record forecasting among the growing uses of AI in finance and supply chain, and their case material reports material reductions in forecast error and inventory. Those results are real for the companies concerned. They also come overwhelmingly from large enterprises with deep data, and the firms themselves stress that data and process readiness decide the outcome.

Where machines win

A model is likely to beat unaided judgement when four conditions hold.

Typical cases in a Nepali business:

In these cases the human forecaster's known weaknesses are costly: anchoring on last month, over-reacting to the most recent event, optimism about favoured products, and inconsistency from one week to the next. A model has none of those. It also does not get tired at the three-hundredth item.

Where judgement wins

A model knows only what its history contains. It fails, often silently, when the future differs from the past in a way the data cannot show.

Structural breaks. Nepal has had many: the 2015 earthquake and the blockade that followed, the pandemic, the import restrictions and liquidity squeeze of 2022, the disruption of September 2025. No model trained on the preceding years would have forecast any of them, and for months afterwards its output was worse than useless.

Policy changes. A budget that alters customs duty or VAT treatment. A Nepal Rastra Bank directive on loan-to-value ratios, working capital limits or margin lending. A change in the IT export tax concession. These move demand and cost by step, on a known date, and they are known to people before they are known to data (what the 2083/84 budget means for businesses).

One-off events. A large tender. A new competitor. A key customer in difficulty. A road closed for a season. A strike.

New things. A new product, branch or market has no history. Analogy and judgement are all there is.

Small data. A company with eighteen months of records, or one whose records changed basis when it changed software, cannot support a model.

Strategic questions. Should we enter this market? Will this investment pay? These are not forecasts in the statistical sense. They are judgements about a future the company itself will shape.

Reflexive situations. If everyone uses similar models and acts on them, the pattern changes. Share markets are the obvious case, which is one reason to be wary of anyone selling AI predictions of the NEPSE index.

Nepal's sources of uncertainty

A forecasting approach for a Nepali business should name the particular things that move its numbers.

Two practical points follow. First, any model must work on the Bikram Sambat calendar or carry festival dates explicitly: Dashain falls in different Gregorian weeks each year, and a model that compares October with October will be wrong every year. Second, the last column shows how much of what matters is scenario and judgement. In an economy this exposed to shocks, the point forecast deserves less attention than the range.

The division of labour

The question is never machine or human. It is which does what.

Step 1: a statistical baseline. For each recurring forecast, produce a baseline by a simple, transparent method. Seasonal naive, moving average or exponential smoothing in a spreadsheet is enough to begin. This is the benchmark everything else must beat.

Step 2: adjust by exception, with reasons. People may change the baseline only for a stated, specific reason the history cannot contain: a confirmed order, a price change, a new outlet, a policy change. Record the adjustment and the reason. Small tweaks "because it feels low" are not allowed.

Step 3: scenarios for the breaks. For the handful of uncertainties that would change decisions, build two or three scenarios with explicit assumptions.

Step 4: measure both. Each month, compare the baseline and the adjusted forecast with the outcome. If adjustments are not adding accuracy, stop making them. If the baseline is failing, find out what changed.

This arrangement uses the machine for consistency and the person for information. It also produces, over a year, evidence about whose judgement adds value, which is more useful than any debate.

A ladder of methods

Climb only as far as the decision justifies.

Most Nepali mid-sized companies are on rung 1 or below. Moving to rung 2 or 3 captures the larger part of the available gain. Rung 5 is appropriate for few, and is usually bought as a feature of an inventory, planning or banking system instead of built (build, buy or wait).

Generative AI tools deserve a specific note. A large language model is not a forecasting engine. Asked to predict next quarter's sales from a pasted table, it will produce a confident, fluent answer by a method you cannot inspect. Its legitimate uses in forecasting are different: writing the formulas or code for a statistical model, explaining a method, summarising what drove a variance, and drafting the narrative. Let it help build the model. Do not let it be the model.

The first forecast to build: thirteen weeks of cash

If a company builds one forecast, it should be a rolling thirteen-week cash flow. It addresses the commonest way that profitable Nepali businesses get into trouble (profitable but short of cash), and most of its inputs are known, not predicted.

Receipts. Start from the receivables ledger. For each customer, apply their actual historical payment delay, not the credit term on the invoice. Add forecast sales for later weeks, lagged by the same habits. This is where a model helps: customer payment behaviour is stable and learnable.

Payments. Payables by due date, adjusted for the company's actual payment practice. Payroll on its dates. Rent. Loan instalments and interest from the bank schedules. LC retirements and import duties from the shipment schedule.

Tax. VAT on the 25th. TDS within 25 days of month end. Advance income tax at the end of Poush, Chaitra and Ashadh. SSF and CIT deposits. These are certain in timing and estimable in amount, and they are the items most often forgotten.

One-offs. Capital expenditure, dividends, bonus, festival advances, insurance renewals.

Position. Opening balance, net movement, closing balance against the overdraft limit, week by week.

Update it every week with actuals, roll it forward one week, and record the variance. Within three months the company will know how reliable each line is. Our cash flow calculator gives a simplified starting view, and how much working capital a business needs covers the underlying arithmetic.

Ranges and scenarios

A single number invites false confidence. Two habits correct it.

Forecast in ranges. State a central estimate with a low and high case, set from the forecast's own track record. If monthly sales forecasts have missed by 12 per cent on average, the honest statement is "NPR 4.0 crore, likely between 3.5 and 4.5". Decisions about stock, staffing and borrowing should be tested against the low end.

Build scenarios around named uncertainties. For a trading company in the current year these might be:

For each, decide in advance the trigger that says the scenario is occurring and the action that follows. A scenario without a trigger and an action is an essay.

Lenders think this way. A projection submitted with a loan application is far more persuasive with a downside case and a debt service calculation under it (what to calculate before taking a bank loan).

Measuring a forecast

Few companies measure forecast accuracy at all. Three measures are enough.

Review these monthly, by forecaster where there are several. Treat a forecast miss as information, not as a failing to be punished: when misses are punished, people forecast low and the numbers become useless. Keep the forecast separate from the target. A target is what the company wants. A forecast is what it expects. Confusing the two is the commonest reason forecasts are wrong.

Common mistakes

From ground level to policy level

At ground level, the forecaster's raw material is the company's own record: sales by item and day, stock, collections. If those records are thin, no method will help, and the first project is recording.

At company level, forecasting is a management process: who produces it, who may change it, which decisions it drives, and how its accuracy is reviewed. The board's interest is in the scenarios and the company's resilience under them.

At sector and national level, Nepal's public data is improving but still limits what models can do. Nepal Rastra Bank publishes monthly macroeconomic and financial data, including remittances, credit, deposits and trade. The National Statistics Office publishes national accounts and prices, and the Nepal Tourism Board and immigration authorities publish arrivals. Much else that a business forecaster would want, such as timely sector output, regional consumption or reliable price series by commodity, is sparse or late. The National AI Policy 2082 speaks of national datasets and open data. For business forecasting, timely and granular public statistics would be among the most valuable things the state could provide.

Lenders and regulators are themselves forecasting more formally. Banks apply expected credit loss models under NFRS 9, which require forward-looking estimates, and stress testing is part of supervision. A borrower whose own forecasts are disciplined speaks the same language.

The role of consulting and advisory firms

The large consulting firms have substantial practices in forecasting and planning: demand forecasting for consumer goods and retail, integrated business planning, treasury and cash forecasting, and scenario analysis. They publish useful material on combining statistical forecasts with planner judgement and on organising planning. The major accounting firms build forecasting and expected credit loss models for banks and review the models others build.

A mid-sized company should take two lessons from that body of work and leave the rest. The first is that process and accountability matter more than the algorithm. The second is that forecast accuracy should be measured. The elaborate platforms are for companies with thousands of series.

Beware two offers. One is the vendor who promises accuracy percentages before seeing your data. The other is the financial projection prepared for a bank or investor with no link to how the business is run. A projection is useful when the company can explain every driver and updates it as results come in.

Our own role is at the practical end. We build driver-based financial models and thirteen-week cash forecasts (financial advisory and modelling), prepare projections for lenders (bank loan documentation), and maintain rolling forecasts as part of virtual CFO work. We use statistical methods suited to the size of the data, and we say when the data does not support a model. We do not sell forecasting software or predictions of markets.

The next 90 days

Days 1 to 30. List the forecasts the business makes and the decision each one drives. Pick two: cash for thirteen weeks, and sales for the main product groups. Gather three years of history on the Bikram Sambat calendar. Build a naive baseline for each.

Days 31 to 60. Run the cash forecast weekly and compare with actuals. Add a seasonal method for sales. Introduce the rule that adjustments need a recorded reason. Define two downside scenarios with triggers.

Days 61 to 90. Calculate error and bias for both forecasts. Decide whether adjustments are helping. Present the board with the base case, the range and the scenarios. Link stock ordering and the borrowing plan to the forecast.

Frequently asked questions

*This article is general guidance, not professional advice. Examples and thresholds are illustrative. Forecasts are uncertain by nature and should not be relied on as predictions of results.*