Agribusiness Management · Agri-sector specialisation · also ABM 526 Business Analytics for Agriculture, Applied Data Science for Food and Agribusiness Management (IIMA), GIS for Agribusiness, ERP, Digital Marketing (MANAGE)
Using data, GIS, ERP and digital channels in agribusiness: the data-science workflow, basic machine learning and ICT for development.
In the exams: agribusiness management PhD entrance and fellowship examinations (light)
Business analytics is the disciplined use of data, statistical models and computation to improve business decisions. Digital agribusiness is the application of the same tools, together with the information systems that collect and move the data, to firms and institutions that buy from farmers, sell to farmers, process food, finance agriculture or advise on it. The subject sits at the meeting point of statistics, computing and management. A seed company forecasting district demand, a dairy cooperative tracking milk quality by collection centre, a lender scoring loan applications from smallholders and a food brand testing two versions of an advertisement are all doing business analytics.
This manual is written for a Master's student in agribusiness management who is not a computer scientist. It follows the units taught in agribusiness MBA programmes worldwide: the data-science workflow and the management of data; regression and classification; forecasting; linear discriminant analysis; principal component analysis; the basics of deep learning; geographic information systems; enterprise resource planning; digital marketing and brand building; and information and communication technology for development, including digital agriculture platforms, electronic markets, agricultural fintech and data governance. Every method is explained with an agribusiness example, and the standard formulas are stated so the usual quantities can be computed by hand.
Analytics is examined lightly in the Agri-Business Management paper of India's ICAR AICE-JRF/SRF PhD entrance (mostly as research-methods and multivariate-technique questions), more heavily inside the coursework of agribusiness MBA and MSc programmes in India, Europe, North America, Australia and elsewhere, and in practice it is tested every day by employers. The questions that recur are conceptual (what does PCA do, what does LDA assume, what is overfitting, what does an ERP system integrate), short numericals (R-squared, a logistic probability, a confusion matrix, NDVI, click-through rate, an A/B test) and judgement (which method fits this problem, what can go wrong with this data). Laws on data protection and on electronic markets differ by country; where they do, this manual says so rather than picking one country's rule.
Analytics is usually divided into three kinds by the question it answers. Descriptive analytics answers 'what happened?' It summarises past data with counts, averages, percentages, trends and charts: tonnes procured per collection centre last month, the share of dealer sales that came from the top ten dealers, average days taken to pay farmers. Diagnostic analysis, sometimes listed as a fourth kind, asks 'why did it happen?' by drilling down and comparing groups. Predictive analytics answers 'what is likely to happen?' It fits a model to past data and applies it to new cases: next season's fertiliser demand by district, the probability that a borrower will default, the expected yield of a field from satellite readings. Prescriptive analytics answers 'what should we do?' It combines a prediction with an objective and constraints and recommends an action: how much of each product to stock, which trucks should go to which villages, what price to offer. Prescriptive methods include linear and integer programming, simulation and decision rules such as the newsvendor critical ratio, which sets the order quantity where the probability of selling out equals underage cost divided by the sum of underage and overage costs. The three kinds build on each other. A prescriptive model is only as good as the prediction inside it, and a prediction is only as good as the description of the data it was trained on. Most of the value in a typical agribusiness comes first from reliable descriptive reporting, because many firms still cannot answer simple questions such as stock by location on a given day.
Analytics projects fail more often from a badly framed question or poor data than from a weak algorithm. A widely taught process model, CRISP-DM (Cross-Industry Standard Process for Data Mining), has six phases. Business understanding states the decision to be improved, who will act on the result and how success will be measured. Data understanding collects the data, describes it and checks its quality. Data preparation cleans, joins, transforms and selects variables; in practice it takes most of the project time. Modelling chooses and fits methods. Evaluation checks whether the model meets the business objective, not only a statistical score. Deployment puts the model into use, whether as a report, a dashboard, a scoring rule inside a loan system or an automated reorder. The process is iterative: evaluation often sends the team back to business understanding. Two habits make the workflow sound. First, define the target and the unit of analysis exactly: is the case a farmer, a field, a loan or a transaction, and is the target known at the moment the decision must be taken? Using information that would not be available at decision time, called leakage, produces models that look excellent in testing and fail in use. Second, hold out data. A model must be judged on data it has not seen, either a separate test set or by k-fold cross-validation, in which the data are split into k parts and each part is used once for testing while the other k−1 parts train the model.
Measurement scale decides which summaries and models are valid. Nominal data are categories without order (crop, district, sales channel). Ordinal data have order but unequal or unknown gaps (grade A, B, C; a five-point satisfaction scale). Interval data have equal gaps but no true zero (temperature in Celsius). Ratio data have a true zero, so ratios are meaningful (tonnes, price, area). Structured data sit in rows and columns; unstructured data include text messages from farmers, images of leaves and voice recordings. Data also differ by source: transaction data from sales and procurement systems, survey data, sensor and machine data, remotely sensed images and third-party data such as weather and market prices. Big data is often described by volume, velocity, variety and veracity; for most agribusinesses veracity, meaning whether the data can be trusted, is the binding constraint. Cross-sectional data observe many units at one time, time-series data observe one unit over time, and panel data observe many units over time.
Operational systems such as a procurement application or a point-of-sale system are built for online transaction processing (OLTP): many small, fast inserts and updates, with the data normalised into many related tables so that each fact is stored once. Analysis needs something different, online analytical processing (OLAP): large read-only queries that summarise across time, products and places. Firms therefore copy data from operational systems into a data warehouse through a process called extract, transform, load (ETL). A warehouse is usually organised as a star schema: a central fact table of measurable events (each procurement transaction, with quantity, price and quality) linked to dimension tables that describe the context (farmer, centre, date, product). A data lake stores raw data of many types, including unstructured files, in their original form, and imposes structure only when the data are read. A data dictionary records what every field means, its unit, its allowed values and its owner. Master data management keeps one authoritative record for each key entity, such as each farmer, dealer or product, so that the same farmer is not registered three times with three spellings. In agribusiness, poor master data is the most common reason why two reports from the same company disagree.
Data quality is usually judged on six dimensions. Accuracy: does the value match reality (is the recorded fat percentage the true one)? Completeness: are required values present? Consistency: do related values agree across fields and systems (does the farmer's village match the collection centre's district)? Timeliness: is the data available when the decision is taken? Validity: does the value conform to the allowed format and range (a moisture reading of 140% is invalid)? Uniqueness: is each real-world entity recorded once? Each dimension can be measured as a percentage, for example completeness = complete records divided by total records, and tracked over time. Cleaning steps include removing duplicates, standardising units and spellings, handling missing values and checking outliers. Missing values should not be filled blindly: deleting incomplete records can bias results if the missing cases differ systematically (small farmers may be the ones without a phone number), and replacing a missing value with the mean shrinks variability. Outliers should be investigated before removal, because a very high yield may be a typing error or a genuinely excellent farm. The cheapest quality control is at the point of capture: drop-down lists, range checks and automatic weighing reduce errors more than any later cleaning.
A key performance indicator (KPI) is a measure that tracks progress towards a strategic or operational objective. Good KPIs are specific, measurable, achievable, relevant and time-bound (SMART), have a clear owner and a target, and are few. Lagging indicators measure outcomes after the fact (profit, market share, default rate); leading indicators move before outcomes and allow action (dealer stock cover before the sowing season, the share of loans with an overdue instalment, farmer enrolment in a contract scheme). The balanced scorecard of Kaplan and Norton groups measures into four perspectives, financial, customer, internal business process, and learning and growth, so that managers do not steer by financial results alone. Common agribusiness KPIs include on-time-in-full (OTIF) delivery, inventory turnover, days of stock, procurement cost per tonne, capacity utilisation of a processing plant, wastage and shrinkage percentage, farmer retention and collection efficiency on loans. A dashboard displays a small set of KPIs for one audience on one screen. Design rules from data-visualisation practice: show each KPI against its target or previous period, use simple charts (lines for trends, bars for comparisons), avoid decorative three-dimensional effects and pie charts with many slices, and let users drill down from a regional total to a centre. A dashboard that nobody acts on is a cost, not an asset; each KPI should map to a decision someone takes.
Supervised learning fits a model to examples in which the correct answer, the target or label, is known, so that the model can predict the target for new cases. When the target is a number (yield, price, sales volume) the task is regression; when it is a category (default or not, disease or healthy) the task is classification. Unsupervised learning, such as principal component analysis or clustering, has no target and looks for structure. The linear regression model writes the target y as an intercept plus a weighted sum of predictors plus an error: y = b0 + b1x1 + b2x2 + … + e. Ordinary least squares (OLS) chooses the coefficients that minimise the sum of squared residuals. In simple regression with one predictor the slope is b1 = Sxy / Sxx, the sum of cross-products of deviations divided by the sum of squared deviations of x, and the intercept is b0 = ȳ − b1x̄, so the fitted line always passes through the point of means. Each slope is read as the expected change in y for a one-unit change in that x, holding the other predictors constant. If both y and x are in natural logarithms, the slope is an elasticity: a coefficient of 0.35 on log nitrogen means a 1% increase in nitrogen is associated with about a 0.35% increase in yield. A dummy variable, coded 1 or 0, shifts the intercept for a group, such as irrigated farms; with a categorical variable of m categories, m−1 dummies are used to avoid the dummy-variable trap.
R-squared is the share of the variation in y around its mean explained by the model: R² = 1 − SSE/SST, where SSE is the sum of squared residuals and SST the total sum of squares. In simple regression R² equals the square of the correlation coefficient. R² never falls when a predictor is added, so adjusted R² = 1 − (1 − R²)(n − 1)/(n − k − 1), with n observations and k predictors, penalises extra predictors. For prediction, error is reported in the units of y: the mean absolute error (MAE) is the average of absolute errors and the root mean squared error (RMSE) is the square root of the average squared error; RMSE punishes large errors more and is never smaller than MAE. The central danger in predictive modelling is overfitting: a flexible model can fit the training data, including its noise, so closely that it predicts new data badly. Training error always falls as complexity rises; test error falls and then rises. This is the bias–variance trade-off: simple models have high bias and low variance, very flexible ones low bias and high variance. Remedies include holding out test data, cross-validation, fewer predictors and regularisation. Ridge regression adds a penalty on the sum of squared coefficients and shrinks them towards zero; the lasso penalises the sum of absolute coefficients and can set some exactly to zero, so it also selects variables. Predictors should be standardised before regularisation so that the penalty treats them equally.
When the target is binary, a linear model can predict probabilities below zero or above one. Logistic regression instead models the log-odds as linear: ln[p/(1 − p)] = b0 + b1x1 + …, so p = 1/(1 + e^−z) where z is the linear score. A coefficient b means that a one-unit rise in x multiplies the odds by e^b, the odds ratio. A classifier turns the probability into a decision with a cut-off, often 0.5 but properly chosen from the costs of the two kinds of error. Results are summarised in a confusion matrix of true positives (TP), false positives (FP), false negatives (FN) and true negatives (TN). Accuracy = (TP + TN)/total; precision = TP/(TP + FP), the share of predicted positives that are real; recall or sensitivity = TP/(TP + FN), the share of real positives caught; specificity = TN/(TN + FP); F1 is the harmonic mean of precision and recall. With imbalanced classes accuracy misleads: if 3% of loans default, a model that predicts 'no default' for everyone is 97% accurate and useless. The receiver operating characteristic (ROC) curve plots recall against the false-positive rate over all cut-offs; the area under it (AUC) is 0.5 for a random classifier and 1 for a perfect one. Other common classifiers are decision trees, which split data by simple rules and are easy to explain, random forests, which average many trees, and k-nearest neighbours.
A forecast is a prediction of a future value of a time series: weekly milk procurement, monthly tractor sales, seasonal demand for a seed variety, the price of an onion in a wholesale market. Agricultural series have strong patterns. The trend is the long-run direction; seasonality is a pattern that repeats over a fixed period such as the crop calendar; cycles are longer, irregular rises and falls such as livestock price cycles; and the remainder is irregular noise. Before modelling, the analyst should plot the series, look for these components, adjust for calendar effects and note structural breaks such as a policy change. Forecasts should be stated with a range, because uncertainty grows with the horizon, and evaluated on data held back from fitting, always forward in time: a model is fitted on earlier periods and tested on later ones, never the reverse. Simple benchmarks must be beaten before a complex model is trusted. The naive forecast sets next period's value equal to this period's; the seasonal naive forecast uses the value from the same season last year. Judgemental forecasts from sales teams are useful but tend to be biased, so they work best combined with a statistical forecast.
A simple moving average forecasts the next period as the mean of the last n observations; a longer window smooths more but reacts more slowly. Simple exponential smoothing gives geometrically declining weights to older data: F(t+1) = αA(t) + (1 − α)F(t), where A(t) is the actual value, F(t) the previous forecast and α, between 0 and 1, the smoothing constant. A high α reacts quickly to change; a low α smooths noise. Holt's method adds a smoothed trend term and the Holt–Winters method adds a seasonal term. In classical decomposition a seasonal index for each season is the average for that season divided by the overall average; an index of 1.25 for the third quarter means that quarter's sales are typically 25% above the average quarter. Dividing actual data by the index gives seasonally adjusted data, and multiplying a trend forecast by the index gives a seasonal forecast. Autoregressive integrated moving average (ARIMA) models, covered in econometrics, model the series through its own lags and past errors after differencing to remove trend. Regression with causal predictors, such as rainfall, area sown and price, is used when those predictors can themselves be known or forecast in time.
Forecast error is actual minus forecast. The mean error, or bias, shows whether forecasts are consistently too high or too low; positive and negative errors cancel in it, so it cannot measure size. The mean absolute deviation (MAD, also called MAE) averages absolute errors. The mean squared error and its root (RMSE) penalise large misses. The mean absolute percentage error (MAPE) averages |error|/actual × 100 and is easy to communicate, but it is undefined when an actual is zero and becomes very large when actuals are small, a real problem for slow-moving spare parts or off-season months. A tracking signal, the running sum of errors divided by MAD, flags a forecast that has become biased. For agribusiness the cost of error is rarely symmetric: running out of seed in the sowing window loses the whole season's sale, while surplus seed can often be carried over at a cost, so the stocking decision should use the forecast distribution and the costs of over- and under-supply, not only the point forecast.
Discriminant analysis classifies cases into known groups from a set of numerical predictors and shows which predictors separate the groups best. In agribusiness it is used to tell adopters from non-adopters of a technology, defaulting from repaying borrowers, members who stay in a cooperative from those who leave, or loyal from switching buyers. Fisher's linear discriminant finds the linear combination of predictors, Z = w1x1 + w2x2 + … + wpxp, called the discriminant function, that maximises the ratio of between-group variance to within-group variance of Z. In words, it finds the direction along which the group means are as far apart as possible relative to the spread inside each group. With g groups and p predictors, the number of discriminant functions is the smaller of g − 1 and p, so two groups always give one function. For two groups with equal prior probabilities and equal misclassification costs, a case is assigned by comparing its discriminant score with the cut-off score halfway between the two group centroids, the mean scores of each group. If priors or costs differ, the cut-off moves towards the centroid of the group that is rarer or less costly to misclassify.
Linear discriminant analysis (LDA) assumes that within each group the predictors are multivariate normal and that the groups share a common covariance matrix. Box's M test checks equality of covariance matrices; if they clearly differ, quadratic discriminant analysis (QDA) is the alternative. LDA is also sensitive to outliers and to strong multicollinearity among predictors, and it needs a reasonable number of cases per predictor. Standard output includes: Wilks' lambda, the ratio of within-group to total variation, which ranges from 0 to 1 with smaller values meaning better separation; the eigenvalue of each function, the ratio of between-group to within-group sums of squares; the canonical correlation, the correlation between the discriminant score and group membership; standardised coefficients, which show each predictor's relative contribution; and the structure matrix, the correlations between each predictor and the function, which are preferred for interpretation when predictors are correlated. Classification accuracy is shown in a classification matrix and should be judged against chance: with two equal groups chance is 50%, while the proportional chance criterion is the sum of squared group proportions. Accuracy should be checked on a holdout sample or by leave-one-out cross-validation, because accuracy on the data used to fit the function is optimistic.
Both LDA and logistic regression classify cases into groups. Logistic regression makes fewer assumptions: it does not need normal predictors or equal covariances, it handles categorical predictors through dummies easily and it gives probabilities and odds ratios that managers can interpret. LDA is more efficient, giving more precise estimates from the same data, when its assumptions hold, extends naturally to more than two groups and gives a low-dimensional view of group separation. For these reasons logistic regression has become the more common choice for credit scoring and adoption studies with mixed predictors, while discriminant analysis remains standard in textbook multivariate analysis, marketing segmentation studies and older credit-scoring work. Altman's Z-score for predicting corporate bankruptcy is a well-known application of discriminant analysis to financial ratios; its original coefficients were estimated on manufacturing firms in the United States and should not be applied to farm businesses or other countries without re-estimation.
Principal component analysis (PCA) is an unsupervised method that replaces a set of correlated variables with a smaller set of uncorrelated variables, the principal components, that keep as much of the original variation as possible. Each component is a weighted sum of the original variables. The first component is the direction of greatest variance in the data; the second is the direction of greatest remaining variance that is uncorrelated with (orthogonal to) the first, and so on. Mathematically the components are the eigenvectors of the covariance or correlation matrix, and each eigenvalue is the variance of its component. With p variables there are p components, and the eigenvalues sum to the total variance; when PCA is run on the correlation matrix, that is on standardised variables, the total variance equals p. The share of variance explained by a component is its eigenvalue divided by the total. Agribusiness uses include building a single index from many correlated indicators (a farm mechanisation index, a market access index, a wealth or asset index from household surveys), reducing many correlated soil or weather variables before regression so that multicollinearity disappears, and compressing spectral bands of satellite images.
Variables measured in different units must be standardised, which means running PCA on the correlation matrix; otherwise the variable with the largest variance, for example income in currency units against family size in persons, dominates the first component purely because of its units. Common rules for how many components to keep are: the Kaiser criterion, keep components with an eigenvalue greater than 1 when using the correlation matrix, because such a component explains more than one original variable; the scree plot, keep the components before the point where the plot of eigenvalues levels off, the 'elbow'; and a cumulative variance target, such as 70–80%, chosen for the purpose. Loadings, the correlations between the original variables and a component, are used to name it: if irrigation share, tractor ownership and fertiliser use all load highly on the first component, it can be read as an 'input intensity' dimension. A variable's communality is the sum of its squared loadings on the retained components and is the share of its variance those components reproduce. Component scores for each case can then be used as variables in regression or clustering. Limitations: components are linear, they may be hard to interpret, and the components with the most variance are not necessarily the most useful for predicting a particular target.
PCA is often confused with exploratory factor analysis. PCA is a data-reduction technique that rearranges total variance into components; it has no model of underlying causes. Factor analysis assumes that observed variables are driven by a smaller number of unobserved latent factors plus unique error, and it analyses only the shared (common) variance. Researchers measuring attitudes, such as farmers' risk perception from a battery of survey statements, use factor analysis and often rotate the solution (varimax is the common orthogonal rotation) so that each variable loads mainly on one factor. Before either method, the Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy and Bartlett's test of sphericity check whether the correlations are strong enough to be worth reducing; a KMO below about 0.5 is usually taken as unacceptable. Because the terms are used loosely in published work, a reader should check whether a study really fitted a factor model or simply ran PCA.
Deep learning is the part of machine learning that uses artificial neural networks with several layers. An artificial neuron computes a weighted sum of its inputs plus a bias, z = w1x1 + … + wnxn + b, and passes it through a non-linear activation function. Common activations are the sigmoid, 1/(1 + e^−z), which maps to values between 0 and 1; the hyperbolic tangent, between −1 and 1; and the rectified linear unit (ReLU), max(0, z), which is the usual default in hidden layers because it is cheap and reduces the vanishing-gradient problem. For classification into several classes the output layer uses softmax, which turns scores into probabilities summing to 1. A network arranges neurons in an input layer, one or more hidden layers and an output layer; 'deep' means many hidden layers. Without non-linear activations, any stack of layers would collapse into a single linear model. In a fully connected (dense) layer with n inputs and m neurons there are n × m weights plus m biases. Training adjusts the weights to minimise a loss function, mean squared error for regression or cross-entropy for classification. Backpropagation computes the gradient of the loss with respect to every weight by the chain rule, and gradient descent moves each weight a small step against its gradient: w(new) = w − η × gradient, where η is the learning rate. Training runs over the data in mini-batches; one full pass over the training data is an epoch.
Convolutional neural networks (CNNs) are designed for images. A convolutional layer slides small filters across the image to detect local patterns such as edges and spots; pooling layers reduce the size of the representation; later layers combine patterns into features such as lesions or leaf shape. With an input of width N, a filter of width F, stride S and no padding, the output width is (N − F)/S + 1. CNNs are used to recognise plant diseases and pests from phone photographs, grade fruit and grain on sorting lines, count animals or plants in drone images and classify crops in satellite images. Recurrent networks and their long short-term memory (LSTM) variant process sequences and have been applied to time series such as prices and sensor readings; transformer models, which rely on an attention mechanism, now dominate language tasks such as answering farmers' text queries. Transfer learning reuses a network pre-trained on a large general image collection and fine-tunes it on a small agricultural dataset, which is how most practical crop-image models are built.
Deep networks have very many parameters and overfit easily, so they need large labelled datasets and regularisation: dropout, which randomly switches off neurons during training; early stopping, which stops training when validation error starts to rise; weight penalties; and data augmentation, which creates extra training images by rotating, flipping or changing brightness. A model trained on clean images taken in a laboratory or on one region's varieties often performs much worse on field photographs from another region, a problem called dataset shift; it must be tested on data from the conditions where it will be used. Deep models are hard to explain, which matters when the output affects a loan or an insurance claim. They also demand computing power and energy. For tabular business data with a few dozen columns, such as a loan file or a sales history, well-tuned regression, tree-based models and gradient boosting often match or beat neural networks and are easier to explain, so deep learning is not the default choice for every problem.
A geographic information system (GIS) captures, stores, analyses and displays data linked to locations on the earth. Because almost every agribusiness decision has a place, where to put a collection centre, which villages a dealer serves, which fields are under a crop, GIS is one of the most practical analytics tools in the sector. Spatial data come in two models. Vector data represent features as points (a warehouse, a well, a farmer's homestead), lines (roads, canals) and polygons (field boundaries, districts), each with a table of attributes. Raster data represent space as a grid of cells, each holding a value, as in satellite images, elevation models and rainfall surfaces; the cell size is the spatial resolution, and a 10 m cell covers 100 square metres. Locations are described in a coordinate reference system: geographic coordinates in latitude and longitude on a datum such as WGS 84, or a projected system such as Universal Transverse Mercator (UTM) in metres. Layers in different systems must be brought into the same system before analysis, or they will not line up. Map scale is the ratio of map distance to ground distance; at 1:50,000, 1 cm on the map is 500 m on the ground. Global navigation satellite systems (GNSS), of which GPS is one, supply positions for field mapping and vehicle tracking.
Core GIS operations include: geocoding, turning addresses or village names into coordinates; buffering, drawing a zone of fixed distance around features, such as a 15 km catchment around a processing plant; overlay, combining layers to find where conditions coincide, such as land that is irrigated, under the target crop and within the catchment; network analysis, finding shortest or cheapest routes along roads and allocating villages to the nearest centre by travel time rather than straight-line distance; interpolation, estimating values between sampled points, such as soil nutrient maps from soil samples; and spatial query. Applications in agribusiness include site selection for warehouses, cold stores and procurement centres; territory design for input sales teams; estimating market potential as cropped area within a dealer's reach multiplied by input use per hectare and expected adoption; route planning for milk collection; mapping supply risk from flood or drought zones; and traceability for deforestation-free sourcing, where buyers must show the location of the plots that supplied a commodity, a requirement that differs by importing market and is changing quickly. Thematic maps display results: choropleth maps shade areas by a value and should use rates (per hectare, per household) rather than raw counts, because large districts otherwise look important simply because they are large.
Remote sensing collects information from satellites, aircraft and drones. Healthy vegetation reflects strongly in the near-infrared (NIR) and absorbs red light for photosynthesis. The Normalised Difference Vegetation Index, NDVI = (NIR − Red)/(NIR + Red), ranges from −1 to +1; dense green vegetation gives high positive values, bare soil low positive values and water usually negative values. Time series of NDVI and similar indices are used to map crop area, monitor crop condition, estimate yields, trigger index insurance and verify that a field financed for a crop was actually sown. Optical images are blocked by cloud, which matters in monsoon and humid-tropical seasons; synthetic aperture radar penetrates cloud and is used for crop and flood mapping. Precision agriculture uses GNSS, yield monitors, soil sampling and sensors to manage variation within fields, for example variable-rate application of fertiliser by management zone. Satellite estimates must be calibrated and validated with ground observations, often called ground truth, before they are used for money decisions.
Enterprise resource planning (ERP) is integrated software that runs a firm's core processes on one shared database. Typical modules are finance and accounting, procurement and purchasing, inventory and warehouse management, production planning, sales and distribution, human resources and payroll, and often quality management, maintenance and customer relationship management. The defining feature is integration: when a procurement officer records a delivery of grain at a collection centre, the same transaction updates stock, creates the liability to pay the farmer in accounts payable, posts to the general ledger and becomes visible to planners, without anyone re-entering it. Before ERP, firms ran separate departmental systems, often called 'islands of automation', that disagreed with each other and had to be reconciled by hand. ERP grew from material requirements planning (MRP), which computes component needs from a production schedule and bill of materials, and its extension, manufacturing resource planning (MRP II). ERP systems are built around standard business processes, sometimes called best practices; firms either adapt their processes to the software or customise the software, and heavy customisation raises cost, risk and the difficulty of later upgrades. Systems can be installed on the firm's own servers or used as cloud software paid by subscription, which lowers upfront cost and has made ERP affordable for medium-sized processors, cooperatives and producer companies.
Agribusiness puts special demands on ERP. Raw material is biological and variable, so the system must record quality attributes (moisture, fat and solids-not-fat in milk, sugar content in cane, grade) and pay by formula. Supply is seasonal and comes from thousands of small suppliers, so farmer master data, advances, input credit and recovery against produce must be handled. Products have shelf lives, so batch and lot tracking with first-expired-first-out picking is essential. Food safety rules require traceability one step back and one step forward, meaning the firm can identify which supplier lots went into which finished batches and which customers received them; a good ERP can run a mock recall in minutes rather than days. Catch-weight items, such as meat sold by actual weight, and co-products, such as bran and husk from rice milling, also need support. Integrated data then feed dashboards on procurement cost, yield from raw material to finished product, plant utilisation and working capital.
ERP projects are expensive and have a high rate of overruns and partial failure. Critical success factors repeatedly found in studies are top-management support, a clear business case, business process re-engineering before or with the software rather than automating the old process, a strong project team that includes users, good data migration and cleaning, user training and change management. Implementation can be big bang, with all modules switched on at once (faster and cheaper if it works, but risky), phased by module or site (slower but each step is smaller), or run in parallel with the old system for a period (safer but costly in double work). Total cost of ownership includes licences or subscriptions, hardware or cloud fees, implementation consultants, customisation, data migration, training, internal staff time and years of support and upgrades; licences are often a minority of the total. Benefits include lower inventory and working capital, faster financial closing, fewer errors, better traceability and one version of the truth for decisions. Many benefits are intangible or appear only after a learning dip in the first months after go-live, so post-implementation review should compare KPIs before and after over a sufficient period.
A brand is a name, term, sign or design, or a combination of them, that identifies a seller's goods and distinguishes them from competitors. Brand equity is the value a brand adds to a product; in the customer-based view of Keller it comes from brand awareness and the strength, favourability and uniqueness of brand associations. In agribusiness, brand building ranges from consumer food brands to farmer-facing brands of seed, crop protection and animal feed, where trust is built through visible field results, dealers and word of mouth, and to collective marks and geographical indications that protect regional products, whose legal form differs by country. Digital marketing uses digital channels to reach customers. Owned media are channels the firm controls (its website, app, messaging lists); paid media are bought (search and social advertising, display banners); earned media are what others say (shares, reviews, press coverage). Main tools are search engine optimisation (SEO), improving organic ranking in search results; search engine marketing (SEM), paid listings usually priced per click; social media marketing; content marketing, useful material such as crop advisories and recipes that attracts customers; email and mobile messaging; influencer and farmer-ambassador programmes; and e-commerce, both business-to-consumer and business-to-business. In rural markets, short video, voice and messaging in local languages often reach farmers better than text-heavy channels.
The marketing funnel describes the path from awareness to interest, consideration, purchase and loyalty, and digital metrics are tied to its stages. Impressions count how often an ad is shown. Click-through rate (CTR) = clicks/impressions. Cost per click (CPC) = spend/clicks. Cost per mille (CPM) = spend/impressions × 1,000, the cost of a thousand impressions. Conversion rate = conversions/clicks (or visitors). Customer acquisition cost (CAC) = marketing spend/new customers acquired. Return on ad spend (ROAS) = revenue attributed to ads/ad spend; it uses revenue, not profit, so a ROAS above 1 can still lose money if margins are thin. Customer lifetime value (CLV) is the present value of the margins a customer will bring over the relationship; with a constant annual margin m, retention rate r and discount rate d over an infinite horizon, one common textbook form is CLV = m × r/(1 + d − r), with margins received at the end of each year the customer is retained. A business is sustainable only if CLV comfortably exceeds CAC. Attribution, deciding which touchpoint earned credit for a sale, is difficult: last-click attribution gives all credit to the final click and undervalues the awareness channels that started the journey.
An A/B test, or online controlled experiment, randomly assigns users to a control version (A) and a treatment version (B) of a page, message or offer and compares a pre-chosen metric. Randomisation makes the groups comparable, so a difference larger than chance can be attributed to the change, which observational comparisons cannot do. For conversion rates the standard analysis is a two-proportion z-test: z = (p_B − p_A)/√[p̂(1 − p̂)(1/n_A + 1/n_B)], where p̂ is the pooled conversion rate; at the 5% significance level with a two-sided test, |z| > 1.96 is significant. Sample size must be fixed before the test from the baseline rate, the minimum detectable effect, the significance level and the power (usually 80%); detecting a small lift on a low baseline needs thousands of users per arm. Common errors: peeking at results daily and stopping as soon as the difference looks significant, which inflates false positives; testing many metrics or variants and reporting the one that 'won'; running the test for less than a full business cycle, ignoring weekly patterns; and a sample ratio mismatch, where the split between arms differs from the design, which signals a bug. Statistical significance is not business significance: a lift must also be large enough to matter. Where randomising individuals is impossible, as in a village demonstration programme, randomising villages or regions is possible but requires larger samples.
Information and communication technology for development (ICT4D) studies how phones, the internet, radio, video and data systems can reduce poverty and improve services. In agriculture the core economic argument is information asymmetry and search cost. When farmers and traders do not know prices in other markets, the same produce sells at very different prices in nearby markets and farmers have weak bargaining power. Well-known studies found that the spread of mobile phones reduced price dispersion across markets, in Jensen's study of fishermen in Kerala, India, and in Aker's study of grain markets in Niger, because sellers and traders could compare prices before choosing where to sell. ICT services in agriculture include price and weather information; advisory services through voice, text, video and call centres, often linked to public extension; digital farmer registries and land records; input subsidy delivery through electronic vouchers or direct transfers; and traceability and certification systems. The digital divide limits impact: access, cost, electricity, literacy, language and gender gaps mean that women, older farmers and remote households are reached less, so programmes that rely only on smartphones can widen inequality. Evaluations find that information alone helps most when farmers can act on it, for example when they have a market to switch to or the inputs that an advisory recommends.
A digital platform is a two-sided or multi-sided market that connects groups, such as farmers and buyers, and gains value from network effects: more buyers attract more sellers and vice versa. Platform types in agriculture include e-marketplaces for produce, input e-commerce, equipment rental and services platforms, and farm management software. Electronic agricultural markets link physical wholesale markets or warehouses through online bidding, aiming at wider buyer participation, transparent price discovery and faster payment; their results depend on assaying and grading, logistics, dispute settlement and payment systems, and on whether traders from outside the local market actually bid. Warehouse receipt systems let a farmer store graded produce in an accredited warehouse and use the receipt, increasingly electronic, as collateral or for sale. Agricultural fintech covers mobile money and digital payments, digital credit scored with alternative data such as transaction histories, satellite crop records and repayment of input credit; index or parametric insurance that pays on a measured weather or satellite index instead of field loss assessment, which lowers cost but creates basis risk, the gap between the index and the farmer's actual loss; and digital savings and supply-chain finance. Credit risk is still assessed with the standard expected-loss formula, EL = probability of default × loss given default × exposure at default.
Data governance is the set of roles, policies and processes that decide who may collect, access, use and share data, and how its quality and security are maintained. Farm data raise particular questions: who owns field-level yield and machine data, whether farmers can move their data between service providers, and whether aggregated data give platform owners market power or trading advantages. In several regions industry and farmer organisations have adopted voluntary codes of practice on agricultural data built on consent, transparency, purpose limitation and the farmer's control over sharing. Personal data protection laws exist in many countries and differ in scope and enforcement. Principles that recur across them include lawful and transparent processing, purpose limitation, data minimisation, accuracy, storage limitation, security, and accountability of the organisation that controls the data; the European Union's General Data Protection Regulation is the most widely cited example. Practical safeguards include collecting only what is needed, informed consent in a language the farmer understands, role-based access, encryption, anonymisation or pseudonymisation before analysis, and retention limits. Algorithms used for credit, insurance or targeting of subsidies can encode bias from their training data, for example scoring women or tenant farmers lower because they lacked past formal credit, so models should be tested for unfair impact on groups, explained to the people affected and subject to human review and appeal.
A student meets this subject first in a project or thesis: cleaning a farmer survey, building an asset or market-access index by PCA, running a logistic or discriminant model of technology adoption, or forecasting arrivals in a wholesale market. The practical tasks are to state the decision and the target before touching data, check completeness and duplicates, standardise variables before PCA, hold back test data, report a confusion matrix rather than accuracy alone and present results in a simple dashboard. For entrance and fellowship exams the student should be able to explain each technique in two sentences, state its assumptions and compute the standard quantities by hand: R-squared, a logistic probability, precision and recall, an exponential-smoothing forecast, variance explained, NDVI, CTR and CAC.
Five years into a seed, feed, dairy, food or lending business, the analyst owns the numbers that managers use. Typical work: a district-level demand forecast for the sales season with a seasonal index and an error report; a dealer or territory dashboard with stock cover and collection KPIs; a credit scorecard for farmer loans validated on a holdout sample and checked for bias; a GIS catchment study for a new collection centre; digital campaigns judged by CAC and conversions rather than likes, with proper A/B tests. The practitioner also becomes the person who argues for fixing data at the point of capture and who explains to managers why a model with 97% accuracy may be useless.
A senior manager with 25 years of experience decides what to build and what to buy. Decisions include whether to implement or replace an ERP system and how (phased or big bang), how much to customise, which KPIs the board sees, whether to join or build a digital platform, how to price data partnerships with input companies, lenders and insurers, and how to meet data-protection law in each country of operation. The senior manager's main analytics job is governance: setting data ownership, quality standards, model approval and review, and making sure that algorithms affecting farmers' credit, insurance or payments are explainable and fair. Experience also teaches scepticism about vendor claims and the value of a clean, trusted master-data foundation before any advanced model.
A researcher with a doctorate and long field experience works on questions the industry cannot answer alone: causal evaluation of digital advisory and e-market interventions with randomised or quasi-experimental designs; whether electronic markets actually widen buyer participation and raise farm-gate prices; how satellite-based yield estimates and alternative-data credit scores perform across regions and for women and tenant farmers; how basis risk can be reduced in index insurance; and what governance of farm data balances innovation, competition and farmers' rights. Such a researcher also builds open benchmark datasets, publishes validation results honestly, including failures, and advises regulators on model risk and data-sharing rules.
This manual is an applied introduction, not a full course in machine learning or GIS. It does not cover clustering methods (k-means, hierarchical) beyond a mention, support vector machines, gradient boosting in detail, time-series ARIMA estimation (see the econometrics manual), Bayesian methods, causal machine learning, natural-language processing, programming in any language or software, cartographic projection mathematics, or the technical configuration of any ERP product. Data-protection, electronic-market and digital-lending rules differ by country and change often; only principles common to many laws are stated, and no country's specific rule, deadline or penalty is given. Platform market sizes, adoption statistics and accuracy figures for commercial crop-image or credit models are left out because published numbers vary widely and are often vendor claims. Worked numbers are invented illustrations, not real data for any firm, crop or country.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.