AI for Finance and Investing - 10 Prompts

Using AI for forecasting and financial analytics isn't just a fun toy for startups. Today artificial intelligence is an indispensable assistant for calculations, checks, and forecasts.

In this article we offer 10 genuinely effective prompts that have already been tested on real examples — financial documents from open sources. For the review we picked a range of models from the GPTunneL aggregator, and here's what we got.

Prompt No. 1

For this analysis we used GPT-5. As the test document we took a telecom company's annual report for 2024 (PDF). We framed the task with the following prompt:

"The attached PDF contains a telecom company's annual report for 2024. Using AI for financial analysis, highlight the key metrics (revenue, profit, EBITDA, debt). Compare with 2023 and draw 3 conclusions for a retail investor in plain language."

Result

First, the model correctly extracted the figures we were interested in:

• revenue – 703.7 billion (in local currency, +16% YoY);
• OIBDA – 246.4 billion (+6%);
• net profit – 49 billion (-10%);
• net debt – 477 billion (1.9 × OIBDA).

The model also explained why profit declined and drew conclusions that would be useful to investors. You can review the full example via this link.

Prompt No. 2

Here we tested Claude Sonnet 4.5 as a financial assistant. For this we used a lightweight csv file specially prepared with GDP and CPI data for 2005–2021. The document had to be compressed, compiled, and reformatted, since the AI couldn't handle the full-size source file. The prompt we gave the model for financial analysis:

"The attached CSV file contains macroeconomic data for a national economy for 2005–2021. Using AI for financial analysis, analyze the dynamics of GDP and CPI, build a forecast for 2022–2023, and explain which factors could cause deviations from the base scenario. Present the result as a table with a short plain-language comment."

Result

Claude did a decent job of generating a logical analysis. The model noted GDP growth up to 2013, a decline following sanctions, and a recovery by 2021. It also provided a forecast for the economic indicators, for example:

• GDP: $1,842 → $1,914 billion;
• CPI: 185.2 → 192.8 (2010 = 100).

However, we didn't get the result right away. Initially the model couldn't "swallow" the large CSV, so the data had to be quite literally spelled out for it. That's a real problem: in practical work, few people want to spend an hour manually preparing a massive dataset.

Conclusion: with the right prompt, Claude Sonnet 4.5 delivers good results, but only if it's given small files.

You can see how the model performed in this dialogue.

Prompt No. 3

As the test model here we picked Perplexity Sonar Pro. The document used in the review was an IFRS report from a global mining company for 2021 (PDF). To test the model, we used this prompt:

"Analyze the IFRS PDF report and use AI for finance to find anomalies: cost spikes, rising debt, margin decline. Build a table of 5 metrics and add recommendations for each."

Result

Perplexity Sonar Pro delivered excellent results. Analyzing the file took the model just 3–5 seconds. Without needing any additional clarification, the model completed the task:

• Extracted the base metrics;
• Laid out the data in a clear table;
• Added a relevant comment and a useful recommendation to each item.

Revenue, operating expenses, CAPEX, debt, and margin were described in detail, and the conclusions matched the actual trends. Importantly, the model didn't just recognize the information — it correctly handled context, taxes, duties, social payments, and capital-investment risks.

To take a closer look at the Perplexity report, follow this link.

Prompt No. 4

The next model in the review was DeepSeek V3.1. To test the model and the effectiveness of the prompt, we used an "Analytical review of the banking sector," published by a national central bank for Q2 2025 (PDF). The task we gave the AI:

"Imagine a scenario: inflation +2%, central bank rate unchanged. Using AI for financial and investment analysis, model the impact on the banking sector: profit, capital, default risks. Present the results as 3 scenarios (optimistic, base, stress)."

Result

DeepSeek V3.1 delivered one of the most detailed and technically solid reports. The model instantly dug into the document's context and described three hypothetical scenarios, explaining the logic behind the impact of an inflation shock. Every scenario DeepSeek V3.1 produced came with concrete figures. For example, the model projects a profit decline of up to -15% in the base case and up to -40% in the stress case.

DeepSeek V3.1's analysis is on par with a review from an industry specialist. The AI didn't just list facts — it also demonstrated the relationships between rates, inflation, risks, and bank behavior. The answer turned out substantive and precise, though possibly too detailed for someone unfamiliar with finance.

You can read the model's report here.

Prompt No. 5

LlaMA 4 Scout was another model we used in the review. We asked it to analyze the same telecom company's annual report for 2024 (PDF). Here's the prompt we gave it:

"The attached PDF contains a telecom company's annual report for 2024. Using AI for finance and investing, estimate the company's approximate value using the DCF method, but with a simplified algorithm — no bulky calculations. Find the key metrics (revenue, OIBDA, net profit, CAPEX, net debt), estimate free cash flow (FCF), and calculate an approximate EV. If data is missing, indicate where it can be found."

Result

LlaMA 4 Scout completed the task at an acceptable level. The model correctly understood what was expected and identified which data was missing. It didn't fabricate any figures for the report. For example, it extracted revenue of 184.3 billion (in local currency) and suggested where to look for the missing numbers: the "Financial Overview," "Business Development Investments," and "Debt Load" sections.

LlaMA 4 Scout didn't just guess at the DCF calculation:

• It built the estimate following the logic of a professional analyst;
• Showed the FCF and EV formulas;
• Explained how the discount rate and CAPEX affect the final valuation.

However, the model couldn't extract all the information from the document, which raises some doubt about the accuracy of the estimate. You can learn more from this dialogue.

Prompt No. 6

We tested the next prompt on Grok 4. As the document to process, we used a daily VIX csv file. The task we set for the model:

"The attached CSV contains historical data for the VIX volatility index (Date, VIXOpen, VIXHigh, VIXLow, VIXClose). Using AI for finance, analyze the dynamics of market fear and calm over the last 10 years. Highlight the 5 days with the sharpest spikes, explain their causes, calculate average values, and determine whether the market is currently in a fear phase or a calm phase."

Result

Grok 4 Fast showed it can work both fast and accurately. It took the AI just a few seconds to extract the relevant information and match volatility spikes to real events. Grok 4 flagged the important events of recent years in its table:

• The COVID-19 pandemic (VIX = 75.47);
• Volmageddon 2018;
• The Chinese "Black Monday" of 2015;
• Brexit 2016;
• The autumn sell-off of 2018.

The model didn't just list the dates — it also wrote short comments explaining the causes behind each spike in the "fear index."

Grok then calculated the average VIX value over the last 12 months (15.8 points) and compared it with the historical average (18.2). It also noted that the market is currently in a calm phase with elevated risk appetite. We got a clear, analytically sound, and well-structured answer.

To see for yourself, check out the test chat.

Prompt No. 7

We couldn't skip Gemini 2.5 Pro either. To test it, we uploaded two files: one containing budget execution data from a construction ministry, and another containing budget execution data from a transport ministry. We asked the AI to compare them with the following prompt:

"Compare the budget execution of the construction ministry and the transport ministry using the two CSV files. Show where there was overspending (actual expenses above plan) and where, conversely, funds went unspent. Build a table with three overspending line items and three savings line items, add the execution percentage, and a short comment for each. At the end, write an overall conclusion — which ministry has better budget discipline and which line items should be optimized."

Result

Gemini 2.5 Pro processed both documents fairly quickly and logically. It correctly matched each file to its ministry, handled the terminology accurately, and provided a clear comparison. For example, the model noted that, formally, there was no budget overspending. It also reported the maximum and minimum fund-utilization figures (99.5–99.6%, 42–70%).

The report turned out clear, with a well-structured table and substantive analysis that included recommendations for both ministries. You can evaluate Gemini's work via this link.

Prompt No. 8

We decided to tackle the next task with Mistral Small 3.1. We gave the model a World Bank csv dataset on foreign direct investment into an emerging economy (2004–2023). The AI received this request:

"Use the World Bank CSV file on foreign direct investment (FDI). Analyze the investment dynamics over the last 20 years: when there was growth, decline, and recovery. Show the average investment volume, minimum and maximum values, assess volatility, and draw a short conclusion. At the end, build a table with 5 recommendations on how to reduce risks and improve the structure of investment inflows, in the style of AI for finance and investing."

Result

Mistral Small 3.1 handled the task excellently. It quickly identified periods of investment growth and decline, correctly pinpointed the peaks (2007 and 2012) and the crisis downturns (2008–2009, 2014–2015, and 2020–2023). The model didn't just build a clear timeline — it explained the reasons behind each trend:

• The global economic crisis;
• Sanctions;
• The coronavirus pandemic.

Mistral's main strength is structured analytics with detailed recommendations on how to boost investment inflows: tax incentives, improving the legal framework, using AI for risk forecasting, and so on. For a more detailed look at the report, use this link.

Prompt No. 9

For this review we used Claude Haiku 4.5. To evaluate the model, we uploaded a csv file with averaged domestic trade data for 2020–2024. The prompt we used for this task:

"The file contains averaged domestic trade data for 2020–2024. Using AI as a financial assistant, write a short summary for an investor: where there's growth/decline, what's supporting the local currency and domestic demand, and what the risks are. At the end, give 3–4 conclusions and recommendations by trade segment."

Result

Haiku generated an easy-to-understand summary. The model noted that the number of retail chains grew while the share of retail sales declined. It also pointed out that profit and balance had shrunk. The AI also proposed several recommendations: focus on large national chains, actively develop e-commerce, and pursue import substitution.

Haiku's major drawback is that it can't handle large volumes of data. We had to compile the file manually and simplify the prompt. See a sample report in the corresponding chat.

Prompt No. 10

For the last prompt in our review we decided to test GPT-5 Nano. For this trial we used a csv file with oil prices, a USD/local-currency exchange rate, and a national stock index for 2019–2024. The prompt we gave the model:

"The file contains data on the average price of Brent crude, the USD/local-currency exchange rate, and a national stock index for 2019–2024. Using AI for financial and investment analysis, find how the oil price, the currency rate, and the stock market are related. Calculate correlations, produce a table, and describe the results in three short plain-language paragraphs."

Result

GPT-5 Nano processed the data quickly and accurately, and proposed realistic correlation coefficients:

• Brent ↔ USD/local currency = 0.11 (weak, nearly zero correlation);
• Brent ↔ national stock index = -0.60 (moderate negative correlation).

The model pointed out to investors that in recent years the local currency's exchange rate has become less dependent on oil prices, and that when assessing prospects and risks, other factors should also be considered: sanctions, restrictions, and monetary policy.

One drawback of this model is the difficulty it had processing the large source file. The AI could only analyze a compressed CSV — the original World Bank + GitHub dataset turned out to be too large for it.

Still, you can review the AI's output here.

Prompts for Financial Analysis Using AI

Conclusion

AI is an excellent solution for analyzing financial documents. A model can provide comparative analysis, formulate recommendations, and flag risks. The key is to write a well-crafted prompt that clearly reflects the goals of the analysis.