Algorithmic Trading vs AI Trading: The Real Difference

Algorithmic Trading vs AI Trading: The One-Sentence Difference

Algorithmic trading vs AI trading comes down to who sets the decision rule. In algorithmic trading a human writes the rule and the computer executes it; in AI trading a model learns the rule, or part of it, from data.

An algorithmic system looks like this: if RSI(14) < 30 and price > [VWAP](/learn/vwap-trading-strategy) then buy 100 shares. Every threshold was typed by a person. The same inputs always produce the same output. An AI system looks different: a gradient-boosted classifier trained on 40 features outputs a probability of a 1% move, or a language model reads a headline and returns a bullish or bearish label. The parameters were fitted, not typed.

The two are not opposites. Every AI trading system is also algorithmic, because it still runs as code on a schedule. The reverse is not true. A Forbes Finance Council piece from October 2025 framed the split as control: with rules you can read why the system acted; with a self-learning model you often cannot (Forbes, 2025). That framing holds up, but it hides a spectrum.

The Spectrum From Rules to Agents

Trading automation sits on five rungs. Each rung hands more of the decision to a model and takes more of it away from a human.

The right-hand columns are the ones people skip. The what is AI trading article covers the top three rungs in more depth; the algorithmic trading beginners guide covers the first.

What Makes a System Algorithmic

A rules-based system has three properties. It is deterministic: the same bar history always yields the same order. It is inspectable: you can print the rule and a colleague can check it. And it is cheap to test: a rule with three parameters can be run across a decade of daily bars in seconds.

Here is a concrete rule set, the kind that runs on a broker API in a few hundred lines of Python:

1. Universe: stocks above $5 with 20-day average volume over 1 million shares.
2. Entry: RSI(14) below 30 and the last close above session VWAP.
3. Stop: 1.5 times the 14-day ATR below entry.
4. Target: 2 times the stop distance.
5. Exit: flatten at 3:55 PM ET regardless.

Nothing in that list learns. If it stops working, a human changes a number. When the market shifts, you know which knob to turn. The quant trading guide covers the second rung, where the form is still human-chosen but the numbers come from a fit.

What Makes a System AI

The AI rungs share one trait: some part of the decision was fitted to data rather than written down. That covers three very different technologies.

Machine learning classifiers

A supervised model takes a feature vector (RSI, ATR percentile, relative volume, sector return, and so on) and predicts a label such as "closes above entry within 30 minutes." The model does not know what RSI means. It finds whatever combination of features separated winners from losers in the training window. If that combination was noise, the model still reports high training accuracy.

Reinforcement learning

An RL agent is not told which trades were good. It is given a reward function and a simulator, then learns a policy by trial and error over millions of simulated steps. The catch is the simulator. If fills, slippage, and market impact in the simulator differ from the live book, the agent learns to exploit the simulator. The reinforcement learning for algorithmic trading article works through what a realistic reward function needs.

Large language models

An LLM reads unstructured text, a filing, a transcript, a news headline, and reasons over it. It is the only rung that can consume text at scale without a human building a feature for each phrase. It is also the only rung that arrives pre-loaded with years of market outcomes, which creates the backtest problem covered next.

Why LLM Backtests Before the Training Cutoff Are Invalid

An LLM backtest on dates before the model's training cutoff is not a backtest. The model may already know what happened. This is weight lookahead: the future is baked into the parameters, not leaked through your data pipeline.

The evidence is direct. A December 2025 arXiv paper, "Detecting Lookahead Bias in LLM Forecasts," sent Llama-3.3-70B a query containing only a ticker and a date, no headline, no transcript, and measured how often the model committed to an up or down answer. That recall propensity was materially positive for every year inside the training window and collapsed to roughly zero in 2024, the first year after the model's December 2023 cutoff (arXiv 2512.23847, 2025). The same paper found that the LLM's headline-based return forecasts were most accurate on exactly the firm-dates where recall was strongest, and that the effect disappeared post-cutoff.

Classical lookahead bias is auditable: you read your code and confirm that no future bar reaches a past decision. Weight lookahead is not auditable from the outside. You cannot grep a 70-billion-parameter model for a memorised closing price. The only clean test is to restrict evaluation to dates after the cutoff, which for popular models leaves a window of months, not decades. Any LLM strategy result you see should state the model, the cutoff date, and the evaluation start date. If it does not, treat it as in-sample.

This is why some builders treat the LLM as a multiplier on a rule-based signal rather than as the signal itself. The rules produce candidates using bar data with an auditable clock. The model then adjusts conviction. If the model's contribution turns out to be zero, the rules still stand on their own backtest.

Overfitting Happens in Both Worlds

Rules-based systems overfit too. They just overfit differently. Bailey, Borwein, Lopez de Prado, and Zhu showed in the Notices of the American Mathematical Society that high simulated performance is easy to reach after testing a relatively small number of strategy configurations, and that analysts rarely report how many they tried (Bailey et al., 2014). Their later paper introduced the probability of backtest overfitting, a way to estimate how likely the best in-sample configuration is to disappoint out of sample (Bailey et al., 2017).

The mechanism is the same on every rung. Each configuration you test is a lottery ticket. Enough tickets and one of them wins the in-sample draw by chance.

A hypothetical to make the count concrete:

• A rules-based RSI/VWAP system with three parameters, each tried at 10 values, is 1,000 configurations.
• An ML classifier with 40 features, 5 hyperparameters at 6 values each, and 3 feature-selection methods is 40 × 7,776 × 3, over 900,000 effective trials before any feature engineering.

The rules system is easiest to protect: hold out the last two years, fit on the earlier data, and refit forward in windows. The backtesting trading strategies guide walks through that procedure, and the overfitting glossary entry covers the warning signs. The ML system needs the same discipline plus purged, time-ordered cross-validation so that adjacent bars do not leak across the split. The LLM system needs a post-cutoff window, which the previous section covered.

Regime Shifts and Latency

Two operational differences separate the rungs once a system is live.

Regime shifts

A rules-based system fails loudly. When a trending market turns choppy, an RSI mean-reversion rule starts losing on the same trades it used to win, and the drawdown appears in the journal within days. An ML model fails quietly. Its predicted probabilities still look confident, because the model has no concept of "the world changed." This is concept drift: the relationship between features and outcomes moved, and the model kept the old mapping. The standard remedy is scheduled retraining on recent data plus drift detectors that compare live feature distributions against the training set. Neither is free. Retraining on a short recent window reintroduces the overfitting problem above.

Latency

The rungs also differ by orders of magnitude in decision time.

A scan of 500 tickers every minute is trivial for the first four rungs. For an LLM it is 500 network calls, each taking seconds, which is why LLM-driven systems work on a filtered shortlist rather than a full universe. Tradewink's day-trade pipeline, for example, sets a hard 12-second timeout on each conviction call so that a slow model never stalls the scan loop, and it only scores candidates that already passed the rule-based screener.

Which One Should You Build First

Build the rules-based system first. Add a model only when you have a specific question the rules cannot answer, and only after the rules have a live paper track record you can compare against.

The decision guide:

1. Do you have a hypothesis you can write in one sentence? Build rules. If you cannot state the edge, a model will not find it for you.
2. Do you have fewer than a few thousand labelled trades? Stay with rules or a simple statistical fit. ML classifiers trained on 200 trades learn the 200 trades.
3. Does the edge depend on reading text (filings, transcripts, news)? That is the one job where an LLM adds something rules cannot. Use it as a filter on rule-generated candidates, and evaluate it only on post-cutoff dates.
4. Do you need to act in under a second? Rules or a pre-trained classifier. No LLM.
5. Can you explain a losing trade to yourself afterwards? If the answer must be yes, stay on the top two rungs.

A worked hypothetical

Take a hypothetical trader with a $10,000 paper account, two years of daily bars for 300 liquid stocks, and Python.

Month one: she codes the five-step RSI/VWAP rule set from earlier. The backtest over the first 18 months produces a hypothetical 410 trades, 52% winners, average winner 1.8R, average loser 1.0R. Expectancy per trade is 0.52 × 1.8 minus 0.48 × 1.0, which is 0.456R. On a 1% risk per trade that is about $46 expected per trade on $10,000, before slippage. The held-out final six months produce 120 trades at 49% winners and 0.38R expectancy. Lower, but positive, and she can explain every trade.

Month three: she has 130 live paper trades. That is too few to train a classifier, so she waits and adds one statistical filter instead: skip entries when 20-day realised volatility is in its top decile, a rule she can test with a single parameter.

Month nine: 600 paper trades in the journal. Now she trains a gradient-boosted classifier on the rule's own signals, with features available at signal time, to score each entry 0 to 100. She does not let the model generate trades. It only adjusts size on trades the rules already chose. Walk-forward validation over the 600 trades shows a hypothetical 4-point win-rate lift on the top quartile of scores. Small, but measurable, and if the model degrades she can switch it off and the original rules keep running.

At every stage there was a simpler system to fall back to and a clean backtest to compare against.

How Tradewink Handles This

Tradewink's design follows a rules-first pattern. Rule-based screeners and strategies (momentum, mean reversion, breakout, VWAP, opening-range) generate candidates from bar data. A single language-model call then scores each candidate's conviction from 0 to 100. That score is an additive, capped boost to the composite ranking rather than a multiplier, and it also scales position size rather than creating the trade. In the default configuration, candidates below a conviction floor of 60 are dropped; scores of 80 and above get full size, scores from 65 to 79 get 75% size, and lower passing scores get half size.

The backtest side follows the lookahead rule. Any replay that consults past AI trade reflections must pass an as-of timestamp so that the scorer cannot read lessons from trades that closed after the simulated clock. The LLM weight problem itself cannot be fixed by code, which is one reason the model stays an overlay on rule-based candidates and the AI limitations page lists it plainly. The full pipeline is documented on the architecture page.

All of this is educational material, not financial advice. Rule-based and AI trading systems both lose money in unfavourable regimes, backtests overstate live results, and you can lose more than you expect. Paper-trade any system before risking capital.