AI for financial analysis: a practical working method

What AI for financial analysis actually means

The phrase covers at least five unrelated activities, and confusing them is the source of most disappointment. First, document work: pulling numbers and language out of annual reports, quarterly filings, transcripts, prospectuses and credit agreements. Second, summarisation and comparison: telling you how this quarter differs from the last four, or how one company describes a risk that a competitor omits. Third, structured prediction: statistical or machine-learning models that map historical features to a future outcome, such as a probability that price closes above a level within twenty days. Fourth, ranking: ordering a universe of candidates by a composite of rules and model scores. Fifth, explanation: turning a ranked candidate into prose a human can audit and argue with.

Only the third activity is prediction, and it is the one where expectations run furthest ahead of results. The other four are where the reliable time savings live. A junior analyst spends most of their week on activities one, two and five. That is exactly the part a language model does well, in seconds, and at a cost that rounds to nothing. Treating AI as a very fast, very literal research assistant with no memory and no accountability gets you far closer to useful output than treating it as an oracle.

There is one more framing that matters. Nobody can honestly claim one strategy is more accurate than another; markets are adversarial and the distribution of outcomes shifts. What a tool can do is detect and describe a disruption — a change in volatility, a break of a level that held for months, an unusual options print, a shift in how a company talks about demand — and show you the rules that fired. That is an educational signal about the state of a market, not a promise about the future. Any product or article that offers you accuracy is either careless or selling.

The four layers of an AI analysis stack

Whether you are building this yourself or buying it, the same four layers exist. Knowing which layer a failure came from is most of the debugging.

Layer one: data acquisition

Prices, volumes, corporate actions, fundamentals, filings, transcripts, options chains, macro series. This layer decides the ceiling on everything above it. Three properties matter more than breadth. Timeliness: how stale is the value you are reading, and does the feed tell you? Point-in-time correctness: does the record show what was known then, or has it been silently restated? Completeness of the identifier map: are tickers, exchanges and corporate-action history reconciled, so that a merger or a symbol change does not appear as a 90% crash?

Practical consequence: when a data feed fails, the honest response is to say so, not to fill the gap with a model estimate presented as a fact. A blank is recoverable. A fabricated print is not, because you will size a position on it.

Layer two: deterministic rules

Moving averages, average true range, trend persistence, distance to support and resistance, liquidity floors, earnings-date exclusions, spread width, credit-to-width ratios. These are arithmetic. They always produce the same answer for the same input, and they can be tested exhaustively. Put every constraint you can express as arithmetic in this layer, because it costs nothing and it never hallucinates. If a candidate fails a rule here, no amount of impressive language later should resurrect it.

Layer three: models and ranking

Here you combine features into a score. It might be a linear weighting you chose, a gradient-boosted model trained on historical outcomes, or a language model asked to rank candidates against written criteria. The two requirements are that the score be reproducible from stored inputs, and that its ingredients be visible. A score you cannot decompose is a score you cannot fix when it starts to drift.

Layer four: explanation and interface

The layer humans actually use. Its job is to show the rules that fired, the levels involved, the position size implied, the maximum loss, and the condition that would invalidate the idea. A good interface makes it easy to reject a candidate. Interfaces that only encourage acceptance are marketing, not analysis.

A repeatable nine-step workflow

This is the sequence to run on a single company or a single trade idea. It takes twenty to forty minutes with AI assistance and half a day without.

1. Define the question narrowly. Not “is this a good company” but “did unit volumes grow this quarter excluding the acquisition, and what did management say about pricing for next quarter”. Narrow questions have checkable answers.
2. Fix the data window. Write down the exact filings and dates you will use. This prevents the most common quiet error: mixing a figure from the latest quarter with a comparison from a year-old summary.
3. Extract, with citations. Ask for each number alongside the sentence or table it came from. No citation, no number.
4. Recompute two ratios by hand. Gross margin and one cash-flow measure are usually enough. If both match, the extraction is probably clean. If either is off, stop.
5. Read the language, not just the numbers. Ask for changed wording between this filing and the previous one in the risk factors and outlook sections. Companies rarely announce deterioration; they rephrase it.
6. Run the deterministic screen. Trend, liquidity, volatility, event calendar. This is where most ideas die, and that is the point.
7. Price the trade, if there is one. Use real prints or real quotes. If you only have an estimate, label it an estimate and haircut it against yourself.
8. Write the invalidation level. One sentence: “this thesis is wrong if X”. If you cannot write it, you do not have a thesis, you have a hope.
9. Size from the maximum loss. Never from the expected gain. Then record the whole thing so you can review it later without reconstructing your reasoning from memory.

Notice that AI touches steps three, five and part of six. The rest is discipline that no model supplies for you. That ratio — a minority of the work automated, all of it verified — is what working AI-assisted analysis looks like in practice.

Reading filings and transcripts with AI

This is the highest-value application and the easiest to get wrong. A quarterly filing is long, repetitive, and deliberately smooth. The interesting content is usually a small change in phrasing, a shifted segment definition, a new line item, or a number that only makes sense next to a different number elsewhere in the document.

Ask for differences, not summaries

A summary of a filing tells you what the company wants said. A diff tells you what changed. Feed the current and prior period together and ask specifically: which risk factors are new, which were removed, where the outlook language weakened or strengthened, which segment definitions moved, and where an accounting policy note changed. Removed risk factors are as informative as added ones, and almost nobody reads for them.

Separate extraction from interpretation

Run two passes. The first pass extracts only figures and quotations with locations, nothing else, no adjectives. The second pass interprets that extracted table without access to the original document. Splitting them makes verification tractable: you check pass one against the filing, and you argue with pass two on its logic. Combined passes hide errors inside plausible narrative.

Watch for adjusted figures

Adjusted, pro-forma, constant-currency and organic figures are all legitimate and all easy to blend accidentally. Insist that every extracted number carry its basis. A margin that moved 300 basis points on an adjusted basis and 40 on a reported basis is two different stories, and only one of them is comparable to last year.

Transcripts: read the questions

The prepared remarks are written. The question-and-answer section is not, and it is where hesitation shows. Useful prompts: which questions were asked more than once, which received an answer that did not contain a number, and which topics analysts dropped compared with last quarter. A question repeated three times by different analysts is a topic management does not want to quantify.

Verification that actually catches errors

Pick one extracted figure at random per document and find it yourself. Check the period label. Check the units — thousands versus millions is the single most common silent error. Check that a percentage change was computed on the same basis in both periods. Four checks, under two minutes, and they catch the overwhelming majority of extraction faults.

Ratios, models and the parts you must not delegate

AI is comfortable with arithmetic when the arithmetic is stated, and unreliable when it must decide what to compute. So state it. Give the formula, give the inputs, and ask for the result plus the substitution shown. Anything the model has to infer — which denominator, whether to include leases, whether a one-off is genuinely one-off — should be your decision, written down.

What to delegate

• Pulling line items into a consistent table across periods and peers.
• Computing stated ratios and flagging period-over-period changes above a threshold.
• Unit and currency normalisation, with the conversion shown.
• Finding the sentence that supports or contradicts a specific claim.
• Restating your own thesis back to you so you can hear how thin it sounds.

What to keep

• Choosing the valuation approach. A discounted cash-flow model and a multiple-based comparison answer different questions; the choice encodes your view of the business.
• Every assumption with a growth rate or a discount rate in it. Models will happily produce a terminal value from a number you did not think about.
• Deciding what counts as normalised earnings. This is judgement about the future of a business, not a text-processing task.
• The final position size and the decision to act. Delegating this is how small mistakes become large ones.

One test for whether a step is delegable: could you check the answer in under a minute against a source? If yes, delegate it. If checking the answer requires the same judgement as producing it, you have not saved any work, you have only moved the risk somewhere you cannot see it.

Rules-based screens versus language models

Both belong in the pipeline, in a specific order: deterministic first, language second. The reason is cost of error. A screen that wrongly excludes a candidate costs you one missed idea, which is survivable and invisible. A language model that wrongly includes a candidate costs you a position, which is neither.

What rules do better

Anything countable. Is price above its 200-day exponential moving average? Has the trend persisted across two years rather than two weeks? Is the average weekly range under a threshold? Is there genuine option volume, or a chain that exists only on paper? Is there an earnings date inside the window? Does the credit received justify the width of the spread? Each of these is a line of arithmetic, each is reproducible, and each can be regression-tested so that a change in logic cannot silently break last month's behaviour.

What language models do better

Everything unstructured. Why does this name suddenly appear in several unrelated news items? Is a filing describing a temporary supply problem or a structural one? Does a company's own language about demand contradict the sell-side narrative? Which of forty surviving candidates share a common cause that a purely numerical ranking would miss? These are reading tasks with no closed-form answer, and a model that reads forty documents in a minute changes what one person can cover in a day.

Ordering them

Screen the universe down with arithmetic, hard-fail anything that breaks a rule, and only then spend model calls on the survivors. This also fixes the economics: language-model calls are the expensive part of any research pipeline, and running them on a universe of thousands instead of a shortlist of dozens is how a tool becomes both slow and costly with no gain in quality.

Backtesting without fooling yourself

A backtest is a claim about a counterfactual, which makes it the easiest place in finance to lie to yourself with arithmetic. Five specific traps account for most inflated results.

Look-ahead bias

Using information that was not available at the decision point. Restated fundamentals are the classic case: the figure in the database today is not the figure that was published then. So are index memberships, analyst estimates, and even adjusted close prices after a corporate action. The control is a point-in-time dataset, or at minimum an explicit lag long enough to cover publication delay.

Survivorship bias

Testing on today's listed universe excludes everything that was delisted, acquired or went to zero — which is precisely the tail you needed to measure. Any backtest on a current index membership list is optimistic by construction.

Overfitting

With enough parameters, any historical series can be explained. Symptoms: a strategy with five thresholds that all landed on suspiciously specific values; performance that collapses when a threshold moves 10%; results that depend on a single year. The controls are fewer parameters, sensitivity testing around every threshold, and a genuinely untouched out-of-sample period. If you have looked at the out-of-sample data, it is no longer out-of-sample.

Ignoring costs and fills

Commissions are the small part. The bid-ask spread, slippage on size, and the assumption that you got filled at the midpoint are the large parts — especially in options, where a quoted spread of ten cents on a fifty-cent contract is a fifth of the trade. Assume the worse side of the spread. If the strategy only works at the midpoint, it does not work.

Regime dependence

A premium-selling strategy tested only through a calm uptrend measures the calm uptrend, not the strategy. Segment results by volatility regime and by rate environment, and report the worst segment as loudly as the average. The average is what you hope for; the worst segment is what you have to survive.

A useful discipline: before you look at returns, write down what result would make you abandon the idea. Deciding the failure criterion after seeing the equity curve is not analysis.

Position sizing and expectancy

No amount of analysis quality compensates for sizing that assumes you are right. Two pieces of arithmetic do most of the work.

Expectancy

Expectancy per trade equals win rate multiplied by average win, minus loss rate multiplied by average loss. A method that wins 40% of the time with an average win of 2.5 times the average loss has positive expectancy; a method that wins 80% of the time with an average loss five times the average win does not. This is why headline win rates are close to meaningless on their own, and why premium-selling strategies with attractive win rates need their maximum loss stated in the same breath as their credit.

Sizing from maximum loss

Decide the fraction of capital you will risk on one idea — many retail traders land somewhere between half a percent and two percent — then work backwards. For a defined-risk options spread, the maximum loss is the width minus the credit, times the multiplier, times contracts; solve for contracts. For a stock position, it is the distance to your invalidation level times shares. If the answer is less than one contract or a handful of shares, the correct action is to skip the trade, not to widen the risk until the position feels worth taking.

Correlation is the hidden concentration

Six positions in different tickers that all depend on the same rate path is one position with extra commissions. Before adding an idea, ask what single event would hurt every open position at once. AI is genuinely useful here: ask it to cluster your open book by common driver rather than by sector label, and it will often name the shared exposure faster than you would.

Eleven failure modes and their controls

• Fabricated figures. Control: every number carries a source location, and one is spot-checked per document.
• Stale data presented as live. Control: display the timestamp next to the price, and label estimates as estimates.
• Silent feed failure. Control: fail loudly. A missing history should abort a scan with a visible error, not produce a thinner answer.
• Fluency mistaken for confidence. Control: require the counter-argument in the same output as the conclusion.
• Mixed periods and units. Control: extraction tables include period and unit columns, always.
• Model-priced options. Control: prefer real prints and real quotes; when only a last trade exists, haircut it against yourself and say so.
• Overfitted thresholds. Control: sensitivity test every threshold, and prefer round numbers you can defend.
• Survivorship in evaluation. Control: keep the losers in the record. A journal that quietly drops abandoned ideas is a marketing document.
• Prompt-injected content. Control: treat scraped pages and documents as data, never as instructions, and never let retrieved text change what the pipeline does.
• Anchoring on the first answer. Control: ask the same question twice with the framing reversed, and be suspicious when the answers disagree.
• Automation without a stop. Control: a human approves every order, and every idea carries a written invalidation level before it is sized.

Building an evidence file

The single habit that improves AI-assisted analysis fastest is keeping a per-idea record that a stranger could audit. Six fields are enough: the question, the sources with dates, the extracted figures with locations, the rules that passed and failed, the invalidation level, and the size with the arithmetic behind it. Add the outcome later, including for ideas you never took.

This record does three jobs. It makes review possible — you can see whether your losses came from bad rules, bad execution or bad sizing, which are three different fixes. It makes the method improvable — thresholds can be tuned against recorded outcomes instead of impressions. And it removes the story-telling that memory does for free: nobody remembers accurately why they entered a trade that went against them.

It also gives an adaptive system something to learn from. A loop that adjusts weights based on recorded outcomes is only as good as the honesty of the record, which is another reason to log rejected ideas and not only executed ones.

Confidentiality, records and disclosure

Three practical points that get skipped. First, do not paste material non-public information or client data into a tool whose retention policy you have not read. If in doubt, work from public filings only. Second, if you produce anything that looks like a recommendation for other people, the rules that apply to you depend on your jurisdiction and your relationship to those people — a tool that generates text does not change that. Third, keep the provenance: which model, which version, which prompt, which sources, which date. Regulators and future-you both ask the same question, which is what you knew and when.

And say plainly what the output is. Educational analysis of a possible market disruption, produced by rules plus AI, is a defensible description. Accuracy, guaranteed returns, or a suggestion that one strategy is inherently superior is not.

How Economove applies this

Economove is built as the four-layer stack described above, and deliberately in that order. Market data comes from real feeds with timestamps; when a required price history is unavailable, the scan reports a global data error rather than quietly producing a thinner result. Deterministic rules run first: trend against the 200-day exponential moving average, multi-year trend persistence, average true range for stop and target distances, multi-timeframe support and resistance overlap zones, liquidity floors, earnings-window exclusions, and for premium strategies the relationship between credit received and spread width.

Options candidates are priced from real chain data — live quotes where the data plan allows, otherwise the last actual trade with a conservative haircut and a visible label saying so. Estimates are never dressed up as quotes. Every idea shows the estimated profit next to the maximum loss, the size implied by the budget you set, the minimum return per trade you asked for, and the levels that would invalidate it. The journal keeps the rejected candidates and the reason each gate dropped them, which is what makes the adaptive loop and your own review meaningful.

What it does not do is tell you a strategy is accurate. It shows you where the rules say a market has been disrupted, with the evidence attached, and leaves the decision where it belongs.

Related reading: AI stock analysis explained, choosing an AI tool for financial analysis and weekly credit spread income.

Frequently asked questions