01 / THE BASICS

Algorithmic trading means using programmed instructions to make trading decisions.

A rule might decide what to buy, when to wait, how much to risk, or how to place an order. Some systems are simple. Others use statistical models or machine learning. Algorithmic trading does not necessarily involve AI, and it does not necessarily mean high-frequency trading.

Automation can make a process repeatable and fast. It can also repeat a bad decision very efficiently. The SEC’s overview describes both useful applications and operational risks; the label “algorithmic” is not a quality rating. SEC overview ↗

A rule is a method. An edge is an advantage.

An edge would need to persist after realistic costs and survive new evidence. A working program alone establishes neither.

02 / FOUR DIFFERENT THINGS

A signal is not a completed trade.

  1. ObservationWhat information was available at that moment?
  2. DecisionDid the rule say enter, wait, or leave?
  3. OrderWhat action was requested, at what size and under what limits?
  4. FillWhat actually executed—or what a paper model estimated?

A quote may move before an order arrives. There may not be enough liquidity at the displayed price. An exit can be delayed, partial, or unavailable. A stop is a request triggered by a condition; it does not guarantee the price you will receive.

Waiting can be the correct output. If the required data is stale, a condition fails, or the risk budget is used up, “do nothing” preserves the meaning of the test.

03 / ALGORITHMS, AI & AGENTS

Related ideas. Different jobs.

Algorithmic describes a process carried out by software. AI describes a model that can supply judgments or predictions within that process. Agentic describes a system that can choose its next steps and use tools. These labels overlap; they are not three competing grades of trader.

Three ways software can help with trading
ApproachWho decides what?An illustrative exampleWhat can go wrong?
Rule-based algorithmPeople specify conditions. Code checks them and acts within fixed limits.Request a paper buy only when all six trend checks pass and data is fresh.A clear, repeatable rule can still lose money or rely on bad data.
AI or learned modelA trained model estimates something; surrounding rules decide how to use its output.Estimate the probability of an upward move from past observations, then apply an entry threshold.Overfitting, misleading confidence and patterns that stop holding in new conditions.
Agentic workflowAn AI system chooses steps and permitted tools in response to what it finds.Notice a missing observation, inspect the source, run a diagnostic and propose a separate test.Wrong tool choices, untrusted inputs, compounding errors, added delay and cost.

The model and agent examples are illustrations, not claims about a deployed CouchChange trading model. A learned model can give repeatable outputs when its inputs and settings are fixed; “AI” does not mean the system retrains itself during every trade.

An agent can operate an ordinary algorithm.

Imagine three jobs: a model estimates market conditions, a rule engine checks whether an order is allowed, and an agent investigates exceptions. One system can contain all three. An agent also may do research without placing any orders at all.

“Agent” has no single universal definition. Here we mean an AI system that selects actions and tools as it works, following Anthropic’s distinction between predefined workflows and model-directed processes. Autonomy adds decisions to evaluate; it does not establish a trading advantage. Anthropic: workflows and agents ↗

Automatic is not the same as agentic.

A script running every minute is automated. A chatbot explaining a chart is AI assistance. A system choosing which checks or tools to run is agentic. None of those facts tells us whether its trades are profitable.

A convincing explanation is not a forecast.

A large language model, or LLM, generates language from learned patterns. It can help draft code, summarize documents, propose hypotheses and explain a result. It can also produce confident errors, invented facts or flawed calculations. NIST identifies this failure mode as confabulation. NIST’s generative AI profile ↗

Useful work for AI

  • Suggest questions worth testing.
  • Help write and review software.
  • Organize evidence and find inconsistencies.
  • Translate recorded rule checks into readable explanations.

Claims that still need proof

  • That a proposed pattern predicts future returns.
  • That the data and calculations are correct.
  • That an apparent advantage survives fees and new conditions.
  • That a modeled order could actually execute.

Even a model trained specifically to predict prices faces uncertainty, changing markets and competition. Connecting a chatbot to a wallet does not create better information, execution, or risk management. More autonomy can scale mistakes as well as useful work.

Guaranteed returns, “can’t lose” language and secret AI money machines are warning signs. Regulators have documented fraud built around those promises. CFTC advisory ↗ Joint investor alert ↗

Where CouchChange fits today

AI assists software development, research proposals and retrospective reviews. The current paper trades are decided by registered deterministic rules. Statistical research variants use recorded data; they are not evidence of an LLM discovering reliable profit. Explanations are written after the fact from saved evidence, not presented as hidden model reasoning.

How we would test whether more AI helps

Our proposed standard is a separate, versioned paper experiment: compare a simple rule with a model-assisted version using the same eligible markets, timing, budget and cost assumptions. Record the model version, inputs, tool actions and failed decisions. Include model costs and delays. Test on later data that was not used to choose the method, and keep every variant’s losses visible.

A new model or prompt is a new intervention to evaluate, not permission to rewrite an ongoing experiment. Useful AI may save research time or catch mistakes even when it adds no trading return. Those benefits should be measured separately.

04 / TRY IT YOURSELF

The price went up. Did the trade make money?

This simplified example spends a total of $100 on one buy, including its entry fee, then sells the acquired quantity. Adjust the price move and per-side fee to see the effect.

−10%+10%
0%2%

A 0.05% adverse-price allowance also applies on each side.

Result after modeled costs-$0.70

A 1% price gain still loses money under these assumptions.

Proceeds from the sale
$99.30
If the full $100 had no costs
+$1.00
Price rise needed to break even
1.71%

Teaching example, not a forecast, exchange quote or replay of a CouchChange trade. It omits spread, depth/size impact, network fees, delays, failed orders, taxes and hosting. Real results may be worse. Changing these sliders does not change any study.

05 / READ THE EVIDENCE

Three stages. Three different kinds of confidence.

BACKTEST

What would the rule have done?

Replay historical data. Useful for finding mistakes, but vulnerable to choosing rules that fit the past or accidentally using future information.

FORWARD PAPER TEST

What does it do as new data arrives?

Freeze a rule and observe subsequent outcomes with simulated orders. This reduces some hindsight problems, but fills and costs are still modeled.

LIVE EXECUTION

What happens with real orders?

Real fills introduce operational, liquidity and financial risk. A successful paper test does not automatically justify this step.

Questions to ask of any result

  • What period and account does it cover?
  • Are fees, open losses and failed exits included?
  • Was the rule chosen before seeing these outcomes?
  • How did cash or a simple holding comparison do?
  • Are several “wins” actually the same overlapping market move?
  • What evidence is missing—and what would change the conclusion?

Trying many rules until one looks good is overfitting risk. Choosing only surviving tokens creates survivorship bias. Using information that arrived after the decision creates look-ahead bias. A fresh holdout is later data reserved to test a rule without tuning it to that period.

A high win rate can hide rare, large losses. A large observation count can be thousands of snapshots of the same event. Neither is enough on its own.

06 / A SMALL GLOSSARY

Words you’ll meet in the research.

Spot trading

Buying or selling the asset itself. The current studies do not borrow money or use leverage.

Liquidity

How much can be bought or sold at available prices. A headline market cap does not tell you how much cash can leave a pool.

Spread

The difference between the best displayed buying and selling prices.

Slippage

The difference between an expected price and the price received. A fixed allowance is only a model of this uncertainty.

Realized and open P&L

Realized profit or loss comes from completed sales. Open P&L estimates what remaining holdings are worth relative to their cost.

Drawdown

The decline from a previous account-value peak. A loss limit can react to a drawdown without guaranteeing a maximum loss.

Paper portfolio and event study

A portfolio has a shared cash budget. An event study measures selected outcomes that may overlap without sharing a budget.

Market capitalization

Token price multiplied by supply. It is not money raised or cash available to sellers.

READ FURTHER

Sources behind the introduction.

These sources inform the educational material. They do not endorse CouchChange or validate its strategies.

Algorithm, model-risk and agent definitions reviewed September 14, 2026; investor-alert sources reviewed September 13. The cost example and descriptions of our own studies are CouchChange explanations, not copied trading recommendations.

Put the vocabulary to work.

Read one actual study design and see why its rules sometimes ask the system to wait.

Explore the Slow trend study