What it means
A neural network takes in a set of inputs, such as a customer's income, payment history and account balance. Each input is multiplied by a weight, which is a number showing how important that input is, and the results are added together.
The network passes this total through further layers, and at the end it gives an output, for example the probability that a loan will default. What makes the model useful is that it learns the weights itself.
During training, it is shown thousands or millions of past cases where the answer is known, and it compares its predictions to the real outcomes. It then nudges the weights to reduce the error, and repeats this process until the predictions stop improving.
In business, this means a neural network can find patterns that a human analyst or a simple formula would miss. Banks use them to flag suspicious card payments in milliseconds, insurers use them to estimate claims costs, and retailers use them to forecast what will sell next month.
They are particularly good when the relationships between inputs are complicated and not simply additive. Neural networks are often called black boxes because it can be hard to explain why they produced a particular answer, which is a problem when regulators or customers demand a clear reason for a decision.
They also need a lot of good data, and if the data is biased or poor, the results will be too. This links to the saying Gold in, Gold out.
Managers do not need to build these models, but they should ask sensible questions. What data was used, how was accuracy tested on new cases, and how will the model be monitored once it is live?
A neural network that performed well last year can degrade quickly if customer behaviour or the economy changes. Variants include deep neural networks, which have many layers, and recurrent networks, which handle sequences such as time series of prices.
All of them follow the same basic idea of weighted inputs, layered calculations and learning from error. Simpler models, such as regression, are often still preferred when transparency matters more than the last few points of accuracy.
In practice
Real-world examples.
Example
A credit card issuer trains a neural network on 20 million past transactions. When a customer in Madrid suddenly appears to spend $3,500 in a Singapore electronics shop, the model scores it as high risk. The payment is paused and a text message is sent to confirm.
Example
A supermarket chain feeds sales, weather and holiday data into a network to forecast demand for ice cream. The forecast helps it order the right stock and cuts waste by $400,000 a year. Store managers are shown the forecast in a simple dashboard.
Example
A lender uses a neural network to help rank small business loan applications. Because regulators require explanations, the team pairs it with a simpler model that lists the main reasons behind each decision. Underwriters make the final call on borderline cases.
Formula
Calculation
Neuron output = activation function of (weight 1 x input 1 + weight 2 x input 2 + ... + bias)
Consider a single neuron in a fraud model with two inputs: the transaction amount in thousands of dollars and the number of card payments in the last hour. A payment of $2,000 gives input 1 = 2, and 5 payments gives input 2 = 5. The weights are 0.4 and 0.2, and the bias is -1. The weighted sum = (0.4 x 2) + (0.2 x 5) - 1 = 0.8 + 1.0 - 1 = 0.8. Using a simple activation that returns the larger of the sum and zero, the output is 0.8, which the next layer treats as a moderate fraud signal.Case study
Seen in the real world.
Falconridge Bank is a fictional lender, and this illustrative story shows the value and the limits of neural networks. Its fraud team replaced a rule-based system with a neural network and found that it caught 30% more fraudulent card payments while raising false alarms by only 2%. Annual fraud losses fell from $9,000,000 to $6,500,000.
Several months later, losses started to climb again, because criminals changed tactics and the data the model had learned from was out of date. The team retrained the network on the latest six months of transactions and set up a monthly review of its accuracy. Losses returned to the lower level, and the head of risk added model monitoring to the bank's standing risk report.
Watch out
Common mistakes.
- Thinking a neural network is intelligent in the human sense. It is a statistical model that finds patterns in data and has no understanding of what the data means.
- Trusting the output without testing it on new data. A model that performs well on the data it was trained on can fail on fresh cases.
- Ignoring explainability. In lending and insurance, firms often need to give clear reasons for decisions.
Questions
People also ask.
Is a neural network the same as artificial intelligence?
It is one technique within artificial intelligence and machine learning, and not the only one.
Does a neural network need a lot of data?
Generally yes, and the larger and more complicated the network, the more examples it needs to learn well.
Can it predict share prices?
It can find patterns in historical data, but markets change and are influenced by news, so no model can predict prices reliably.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%