Should Quant Traders Keep Inventing Indicators, Combine Them with Machine Learning, or Start from Raw Data?

Indicators, machine learning, and raw-data models are layers of one research process. The deciding tests are out-of-sample robustness, costs, risk, and whether complexity earns its place.

Layered statistical surfaces connect market data, feature representations, portfolio decisions, and risk controls.

After reading https://insights.wisdomchain.com/the-mathematics-behind-the-worlds-most-successful-quantitative-traders-what-we-can-learn-and-actually-implement/, I was left with a practical question:

Should we continue developing simple but meaningful technical indicators, use machine learning to integrate many indicators, or train models directly from raw market data?

This is not only a technical question. It goes to the heart of quantitative trading:

How should we transform market information into decisions that remain useful outside a backtest?

My conclusion is that these three approaches are not mutually exclusive. They are different layers of the same research process. For most retail traders, the strongest path is neither a single indicator nor an opaque artificial-intelligence model. It is a disciplined combination of:

  1. economically meaningful features,
  2. appropriately simple predictive models,
  3. realistic validation,
  4. risk-aware portfolio construction, and
  5. controlled execution.

Let us examine why.

Key Takeaways

  • Indicators, machine learning, and raw-data models are complementary layers, not mutually exclusive choices.
  • Begin with an economically plausible hypothesis and a small, interpretable baseline; add complexity only when it improves time-based out-of-sample results after realistic costs.
  • Handcrafted features can make market behavior more interpretable, while machine learning can model interactions; neither prevents data snooping or overfitting.
  • End-to-end learning is most defensible when the data are abundant, clean, point-in-time correct, and structurally rich enough to justify its complexity.
  • A useful forecast must survive portfolio construction, execution, and risk controls; prediction accuracy alone is not a trading objective.

1. What is a technical indicator, fundamentally?

Consider the Relative Strength Index, or RSI.

RSI takes a sequence of recent price gains and losses and compresses that history into a number between 0 and 100. It does not create information. It reorganizes information already contained in prices.

The same is true of other familiar indicators:

  • A moving average compresses recent prices into an estimate of trend.
  • MACD compares trends measured over different horizons.
  • Average True Range estimates the magnitude of recent price movements.
  • Bollinger Bands express the position of price relative to a moving average and recent volatility.
  • Volume-based indicators attempt to measure the relationship between participation and price movement.

In machine-learning terminology, technical indicators are handcrafted features.

They transform raw observations:

Xₜ = (Pₜ, Pₜ₋₁, Pₜ₋₂, …, Vₜ, Vₜ₋₁, …)

into a smaller feature vector:

Zₜ = (RSIₜ, Momentumₜ, Volatilityₜ, VolumeShockₜ, …)

The purpose is not merely compression. A good feature attempts to make a potentially useful market behavior more visible.

For example:

  • Momentum features ask whether returns tend to continue.
  • Mean-reversion features ask whether unusually large moves tend to reverse.
  • Volatility features measure the market’s current risk environment.
  • Liquidity features estimate how easily positions can be entered and exited.
  • Breadth features measure whether a market move is broadly supported.
  • Relative-strength features compare one asset with its peers or benchmark.

This gives simple indicators genuine value. They introduce structure, improve interpretability and allow a trader to express a hypothesis precisely.

However, an indicator is not automatically a trading edge.

A broad academic review by Cheol-Ho Park and Scott H. Irwin found mixed evidence. Many studies reported profitable technical strategies, but the authors also identified serious problems, including data snooping, after-the-fact rule selection, inadequate transaction-cost treatment and difficulty estimating risk. Park and Irwin (2007)

That distinction is crucial:

An indicator measures something. A strategy must demonstrate that the measurement helps predict an outcome after costs and risk.

RSI may accurately measure recent relative strength while still failing as a standalone “RSI below 30 means buy” strategy.


2. Why should we continue developing indicators?

Yes, we should continue developing them, but our goal should not be to create another arbitrary chart overlay.

We should develop indicators when they represent a credible source of information.

A useful indicator should answer a question such as:

  • Is the asset trending or oscillating?
  • Is volatility expanding or contracting?
  • Is current volume unusual for this time of day?
  • Is the asset strong relative to its sector?
  • Is the price move confirmed across related assets?
  • Is the market liquid enough to trade economically?
  • Is an apparent opportunity large enough to survive costs?
  • Has the market moved into a different regime?

The best new indicators may not look like traditional indicators at all. They might measure:

  • order-book imbalance,
  • options-implied skew,
  • dispersion among stocks,
  • analyst-estimate revisions,
  • changes in credit spreads,
  • earnings-call sentiment,
  • unusual ETF flows,
  • cross-asset correlation,
  • liquidity deterioration, or
  • macroeconomic surprise.

These are still indicators in the broader sense. They are mathematical representations of potentially relevant information.

A practical test for a new indicator

Before building one, a retail researcher should be able to complete this sentence:

“I expect this variable to contain information because…”

For example:

“I expect volatility-adjusted momentum to be more informative than raw momentum because a 5% move in a low-volatility stock is more unusual than a 5% move in a highly volatile stock.”

That statement creates a hypothesis that can be tested.

Compare it with:

“I tried 500 combinations of indicators, and this one produced the highest Sharpe ratio.”

The second approach is much more likely to discover historical noise.


3. What does machine learning add?

Suppose we have 20 credible features:

  • momentum at several horizons,
  • RSI,
  • distance from moving averages,
  • realized volatility,
  • volume surprise,
  • overnight return,
  • market return,
  • sector return,
  • relative strength,
  • liquidity,
  • valuation,
  • earnings revisions, and
  • perhaps sentiment.

A simple strategy might assign fixed weights:

Signalₜ = 0.4 × Momentumₜ − 0.2 × RSIₜ + 0.2 × VolumeShockₜ − 0.2 × Volatilityₜ

But why should these relationships be linear? Why should the weights remain constant? And why should RSI have the same meaning during a low-volatility uptrend and a high-volatility market decline?

Machine learning can help answer questions such as:

  • Does momentum work differently at different volatility levels?
  • Is volume important only when liquidity is low?
  • Does an oversold signal matter only when the longer-term trend is positive?
  • Do fundamentals and price signals reinforce or contradict each other?
  • Are interactions among features more informative than any feature alone?

A major study by Shihao Gu, Bryan Kelly and Dacheng Xiu compared machine-learning methods for asset-pricing applications. It found that trees and neural networks could improve forecasts by capturing nonlinear interactions. Importantly, the dominant predictive variables still included economically recognizable measures related to momentum, liquidity and volatility. Gu, Kelly, and Xiu (2018)

This is an important result because it suggests that machine learning does not necessarily make traditional market concepts irrelevant.

Instead:

Machine learning can discover when, where and in what combinations familiar signals are useful.

The indicators provide a vocabulary. The model learns how to construct sentences with that vocabulary.


4. Why not simply feed raw data into a neural network?

The third approach is end-to-end learning:

Raw data → Model → Forecast

For example, a model could receive sequences of:

  • open, high, low and close prices,
  • volume,
  • bid and ask prices,
  • order-book updates,
  • financial statements, or
  • text from news and company filings.

The model then attempts to learn useful representations automatically.

This approach is attractive because handcrafted indicators may discard information. If RSI reduces 14 days of price changes to one number, two very different price paths could produce similar RSI values. A sufficiently capable model might distinguish between them.

End-to-end learning has transformed areas such as image recognition and natural-language processing. But financial markets present special difficulties.

The market has a low signal-to-noise ratio

Prices respond to many participants, goals and information sets. The predictable component of short-horizon returns may be extremely small relative to random variation.

The data are not stationary

A relationship learned from one period may weaken or reverse when:

  • interest rates change,
  • market structure changes,
  • participants adopt the same strategy,
  • liquidity changes,
  • regulations change, or
  • the economy enters a different regime.

Financial observations are not truly independent

Today’s return, volatility and liquidity conditions are connected with yesterday’s. Stocks are also connected through sectors, factors, indexes and macroeconomic conditions.

Large models create more opportunities to overfit

More model capacity does not guarantee more usable information. It may simply enable the model to memorize historical accidents.

Retail traders have limited data and infrastructure

A large institution might have:

  • clean point-in-time databases,
  • tick-level market data,
  • corporate-action expertise,
  • alternative datasets,
  • sophisticated computing systems,
  • execution models, and
  • teams dedicated to validation and risk management.

A retail trader usually has significantly less data, less computing power and less ability to estimate market impact. That often makes an interpretable model with robust features more appropriate than a deep neural network trained on limited observations.

The real question is therefore not whether end-to-end learning is theoretically superior. It is whether the researcher has enough clean, representative data to justify the model’s complexity.


5. The apparent choice is actually a hierarchy

Rather than treating indicators, machine learning and raw data as competing philosophies, we can place them in a common research hierarchy:

Layer 1: Raw information

Examples include:

  • prices,
  • volume,
  • order books,
  • fundamentals,
  • options,
  • macroeconomic data,
  • news,
  • filings, and
  • related-market information.

Layer 2: Representation

This includes:

  • returns,
  • RSI,
  • momentum,
  • volatility,
  • valuation ratios,
  • standardized surprises,
  • sentiment scores, and
  • cross-sectional rankings.

Layer 3: Prediction

Models estimate something useful, such as:

  • the probability of a positive return,
  • expected return,
  • expected volatility,
  • downside probability,
  • probability of a trend continuing, or
  • expected trade profitability after costs.

Layer 4: Portfolio construction

Predictions are converted into:

  • position sizes,
  • asset weights,
  • exposure limits,
  • diversification constraints, and
  • hedges.

Layer 5: Execution and control

The strategy determines:

  • order type,
  • trading time,
  • acceptable spread,
  • turnover limits,
  • maximum loss,
  • emergency shutdown rules, and
  • monitoring procedures.

A predictive model is only one component of this system.

A model can be statistically accurate but financially useless if it:

  • trades too frequently,
  • concentrates in correlated positions,
  • takes excessive tail risk,
  • depends on prices unavailable in real time, or
  • produces returns smaller than spreads, commissions, slippage and taxes.

This is why the best prediction model does not automatically produce the best trading strategy.


6. The most important distinction: prediction versus decision

Imagine that an ML model predicts a stock’s probability of rising tomorrow as 51%.

Is that tradable?

Not necessarily.

We still need to know:

  • How large is the expected move?
  • What is the expected loss if the prediction is wrong?
  • What are the spread and commission?
  • How often does the model trade?
  • How correlated is the trade with existing positions?
  • Is 51% unusual relative to the model’s normal output?
  • Does the model remain calibrated in the current regime?

A trading system should optimize a decision, not merely an accuracy score.

A more realistic target is:

Expected net return = expected gross return − transaction costs − slippage − financing costs

The position then needs to be adjusted for risk:

Position size ∝ expected net return / estimated risk

Even this is simplified, but it illustrates the central issue.

A classifier with 55% directional accuracy may be unprofitable if its losing trades are large or its turnover is excessive. A model with less than 50% directional accuracy might still be useful if it correctly identifies a small number of large moves.

For retail traders, this means “prediction accuracy” should rarely be the primary performance measure.

More useful measures include:

  • return after estimated costs,
  • maximum drawdown,
  • volatility,
  • Sharpe or Sortino ratio,
  • profit factor,
  • turnover,
  • exposure,
  • average trade,
  • worst historical period, and
  • consistency across markets and time periods.

7. The greatest common danger: backtest overfitting

Simple indicators and sophisticated ML models share the same enemy: repeated experimentation on the same history.

Suppose a trader tests:

  • 10 RSI periods,
  • 10 entry levels,
  • 10 exit levels,
  • 5 stop-loss rules, and
  • 5 markets.

That is already 25,000 possible combinations.

One of them will probably look excellent by chance.

Machine learning can scale this problem dramatically. A computer can test thousands of variables, transformations, algorithms and hyperparameters. The ability to search more possibilities increases the need for stronger validation.

Research by David H. Bailey, Stephanie Ger, Marcos López de Prado, Alexander Sim and Kesheng Wu shows how a strategy selected as “optimal” from many variations can perform poorly on new data, even when the underlying returns are random. Their framework formalizes how repeated selection can make an in-sample winner look convincing despite little out-of-sample value. Bailey et al. (2014)

This leads to a counterintuitive principle:

The more strategies you test, the less confidence you should place in the best backtest unless you explicitly account for the search process.

8. A practical framework for retail traders

For a retail trader, I would recommend a staged approach.

Stage 1: Define the economic or behavioral hypothesis

Do not begin with an algorithm. Begin with a reason.

Examples:

  • Trends may persist because investors react gradually to information.
  • Short-term reversals may occur after liquidity-driven price pressure.
  • High-volume breakouts may be more reliable than low-volume breakouts.
  • Volatility tends to cluster.
  • Assets with strengthening fundamentals and positive momentum may behave differently from assets with momentum alone.

The hypothesis does not need to be correct. It must be clear enough to be tested and rejected.

Stage 2: Build a small, understandable baseline

Start with a rule or simple statistical model.

For example:

  • Trade only liquid stocks.
  • Require price to be above its long-term moving average.
  • Rank stocks by medium-term momentum.
  • Avoid positions with unusually high volatility.
  • Rebalance weekly or monthly.
  • Apply a fixed portfolio-risk limit.

This baseline tells you whether additional complexity adds value.

If a sophisticated model cannot reliably outperform a simple baseline after costs, the sophistication has not earned its place.

Stage 3: Use indicators as measurements, not commandments

Avoid treating “RSI below 30” as a universal buy signal.

Instead, measure:

  • the actual RSI value,
  • its change,
  • the percentile of RSI relative to that asset’s history,
  • the long-term trend,
  • current volatility,
  • relative volume, and
  • market regime.

Then test whether the meaning of RSI depends on context.

For example, an oversold reading may behave differently in:

  • a long-term uptrend,
  • a long-term downtrend,
  • a high-volatility market, and
  • a low-volatility range.

Stage 4: Add machine learning incrementally

A sensible progression might be:

  1. linear or logistic regression,
  2. regularized regression,
  3. decision trees,
  4. random forests or gradient boosting,
  5. more complex neural networks only if justified.

This progression matters because every increase in model complexity should produce measurable out-of-sample improvement.

For tabular data built from technical, fundamental and cross-sectional features, tree-based models are often a useful intermediate step. They can capture nonlinearities and interactions without requiring the data volume usually associated with deep learning.

Stage 5: Split data by time, not randomly

Ordinary random train-test splitting can leak future market regimes into the training sample.

A more realistic design is:

  • Training period: fit the model.
  • Validation period: select features and hyperparameters.
  • Test period: perform one final evaluation.
  • Paper-trading period: observe the complete process in real time.
  • Small-capital deployment: begin with limited risk.

A walk-forward process is even better:

  1. Train using only information available up to a certain date.
  2. predict the next period,
  3. move forward,
  4. retrain if the strategy requires it, and
  5. repeat.

At every step, the model must behave as if the future is genuinely unknown.

Stage 6: Include realistic frictions

At minimum, test:

  • commissions,
  • bid-ask spread,
  • slippage,
  • delayed execution,
  • turnover,
  • short-borrow costs where relevant,
  • rejected or partially filled orders, and
  • taxes where they materially affect the intended strategy.

A simple robustness test is to double your estimated trading costs. If the strategy immediately becomes unattractive, its margin of safety may be too small.

Stage 7: Test robustness rather than searching for perfection

Ask:

  • Does the strategy work with RSI periods of 12, 14 and 16, or only 14?
  • Does it survive slightly different entry thresholds?
  • Does it work across multiple assets?
  • Does it work in rising, falling and sideways markets?
  • Does it depend on a few exceptional trades?
  • Does performance disappear after a particular year?
  • Does it remain viable when execution is delayed?

A plateau of acceptable parameters is more credible than one spectacular parameter combination surrounded by failures.

Stage 8: Separate signal quality from position sizing

Do not let strong model confidence automatically create an unlimited position.

Use explicit controls such as:

  • maximum percentage per position,
  • maximum sector exposure,
  • maximum total leverage,
  • volatility-scaled position sizing,
  • daily and monthly loss limits,
  • turnover limits, and
  • an emergency stop for abnormal behavior.

Regulatory guidance for professional algorithmic trading emphasizes general risk assessment, software testing, system validation, trading controls and compliance. Retail systems may be smaller, but the engineering principles remain valuable. FINRA

Stage 9: Compare against honest benchmarks

A strategy should be compared with:

  • buy and hold,
  • an appropriate market or factor benchmark,
  • a simple equal-weight portfolio,
  • the simple indicator baseline, and
  • the same strategy without the ML layer.

This reveals where the improvement actually comes from.

Perhaps the ML model does not improve security selection but reduces drawdowns. Perhaps it adds nothing after turnover. Perhaps the real value comes from volatility scaling rather than prediction.

Without such comparisons, every component can appear essential even when only one is doing useful work.

Stage 10: Monitor model decay

Once deployed, track:

  • live versus expected return,
  • live hit rate,
  • prediction distribution,
  • feature distribution,
  • turnover,
  • slippage,
  • drawdown,
  • concentration, and
  • differences between simulated and actual orders.

Do not automatically retrain a model simply because recent performance is weak. First determine whether the cause is:

  • ordinary randomness,
  • data errors,
  • execution problems,
  • feature drift,
  • a regime change, or
  • the disappearance of the underlying effect.

9. Three research paths, and when each makes sense

Path A: Indicator-based rules

Best suited for:

  • limited data,
  • beginning researchers,
  • lower-frequency trading,
  • situations requiring high interpretability, and
  • strategies based on a clear behavioral hypothesis.

Advantages:

  • transparent,
  • easy to debug,
  • relatively inexpensive,
  • less prone to hidden data leakage, and
  • easier to monitor.

Limitations:

  • rigid thresholds,
  • poor handling of interactions,
  • possible oversimplification, and
  • temptation to optimize parameters excessively.

Path B: Machine learning over engineered features

Best suited for:

  • several credible information sources,
  • moderate or large clean datasets,
  • possible nonlinear relationships,
  • cross-sectional ranking, and
  • researchers who can maintain a rigorous validation process.

Advantages:

  • combines many signals,
  • captures interactions,
  • manages redundancy,
  • allows richer conditional relationships, and
  • can estimate probabilities rather than binary rules.

Limitations:

  • easier to overfit,
  • may be harder to interpret,
  • requires careful feature timing,
  • magnifies data-quality problems, and
  • may increase turnover.

Path C: End-to-end learning from raw data

Best suited for:

  • very large and clean datasets,
  • order-book or high-frequency sequences,
  • text or other unstructured information,
  • well-funded research environments, and
  • cases where handcrafted features may discard important structure.

Advantages:

  • automatic representation learning,
  • potential discovery of unfamiliar patterns,
  • ability to process complex sequences or text, and
  • less dependence on traditional indicator definitions.

Limitations:

  • substantial data requirements,
  • high computational cost,
  • low interpretability,
  • severe leakage and overfitting risk,
  • difficult debugging, and
  • challenging live implementation.

For most retail traders, Path B should come after Path A, not before it, and Path C should be attempted only when the nature and quantity of the data justify it.


10. What should retail traders actually build?

A practical retail research architecture could look like this:

Data

Use adjusted daily or hourly data for sufficiently liquid instruments. Make sure the adjustment process correctly handles splits, dividends and corporate actions.

Features

Start with a compact set:

  • one short-term momentum measure,
  • one medium-term momentum measure,
  • distance from a long-term moving average,
  • realized volatility,
  • volume surprise,
  • market-relative return,
  • sector-relative return, and
  • perhaps one fundamental or macro variable.

Avoid creating 100 variations of essentially the same indicator.

Target

Choose a target connected to the intended decision:

  • next-week return,
  • probability of outperforming the market,
  • volatility over the next month, or
  • probability of a drawdown exceeding a threshold.

For many retail traders, predicting relative returns over several days or weeks may be more realistic than predicting the direction of the next minute.

Baseline

Build a simple linear score or ranking model.

ML challenger

Try regularized regression or a small tree-based model using exactly the same information.

Validation

Use walk-forward testing with transaction costs and no future information.

Portfolio

Hold a diversified group of the strongest signals rather than placing the entire portfolio behind one prediction.

Execution

Trade at a specified time using an assumption that could plausibly be achieved live. Do not generate a signal from the day’s closing price and assume execution at that same closing price unless your process can actually do that.

Deployment

Paper trade first, then deploy limited capital. Increase exposure only after live behavior is reasonably consistent with the tested process.


11. What should we optimize?

This may be the most important part of the discussion.

A great deal of quantitative research implicitly optimizes:

How accurately can I predict returns?

But a trader’s actual objective is closer to:

Objective = expected return − costs − risk penalty − instability penalty

That changes the research priorities.

A slightly weaker forecast may be preferable if it is:

  • more stable,
  • cheaper to trade,
  • easier to understand,
  • less concentrated,
  • less sensitive to parameters, and
  • less dependent on a specific market regime.

The most sophisticated model is not necessarily the one with the most parameters. It may be the model that makes the best trade-off among information extraction, uncertainty, cost and risk.


12. My answer to the original question

Should we continue developing simple technical indicators?

Yes, if they express a meaningful hypothesis or measure a market condition more effectively.

We do not need thousands of cosmetically different indicators. We need better representations of momentum, liquidity, volatility, valuation, sentiment, positioning and market structure.

Should we use machine learning to integrate indicators?

Yes, when we have enough clean data and a disciplined validation process.

Machine learning is particularly useful for combining weak signals, modeling nonlinear interactions and determining when a signal is relevant. But it should be tested against a simpler baseline and judged after costs.

Should we build models directly from raw data?

Sometimes, but not by default.

Raw-data models make sense when the data are abundant and complex enough to support representation learning. For most retail traders using daily price and volume data, end-to-end deep learning may introduce more estimation error than usable insight.


The deeper conclusion

The debate is not really:

Technical indicators versus machine learning.

It is:

Human-designed representations versus machine-learned representations, and how to combine them without mistaking noise for knowledge.

Indicators bring economic intuition, compression and interpretability.

Machine learning brings flexible combination, interaction modeling and automation.

Raw-data learning brings the possibility of discovering representations humans did not anticipate.

But none of them eliminates the need for:

  • clean data,
  • a credible hypothesis,
  • honest out-of-sample testing,
  • transaction-cost modeling,
  • diversification,
  • position sizing,
  • execution controls, and
  • continuous monitoring.

For retail traders, the best path is usually:

Simple hypothesis → Meaningful features → Simple baseline → Carefully validated ML → Risk-controlled deployment

The durable advantage is unlikely to come from discovering a magical indicator or downloading a supposedly intelligent trading bot. FINRA has specifically warned investors about unregistered auto-trading providers advertising beginner-friendly, “risk-free” systems, unusually consistent returns or an AI advantage. FINRA

The advantage, if one exists, comes from building a better research process.

A simple indicator is valuable when it measures something real. Machine learning is valuable when it combines information without overfitting it. Raw data are valuable when we have enough evidence to learn from them. But risk management determines whether any of those insights survive contact with the market.

That, to me, is the essential mathematics behind successful quantitative trading: not complexity for its own sake, but the disciplined conversion of uncertain information into controlled decisions.


References

This discussion is educational and describes a research process, not a recommendation to buy or sell any security.

Read the Chinese companion: 指标、机器学习与原始市场数据:量化交易研究应如何选择.