Home / Backtesting Biases / Underfitting
BiasIntermediate

Underfitting

Underfitting is the condition in which a strategy or model is too simple or too constrained to capture the genuine structure in the data, so it performs poorly both in-sample and out-of-sample because it has failed to learn the real signal that is present.

Quick answer: Underfitting is the condition in which a strategy or model is too simple or too constrained to capture the genuine structure in the data, so it performs poorly both in-sample and out-of-sample because it has failed to learn the real signal that is present.

In simple words

Underfitting is bringing a butter knife to carve a detailed sculpture: the tool is too blunt for the job. A model that underfits is so simple that it misses a real, exploitable pattern, so it does badly on both old and new data. It is the opposite mistake to overfitting, and it is the price of being too cautious rather than too greedy.

Purpose

This concept exists because the fear of overfitting can push researchers so far toward simplicity that they miss real edges, so understanding underfitting keeps the pursuit of robustness from becoming a refusal to model anything at all.

Professional explanation

What underfitting is

Underfitting occurs when the model's capacity is below the complexity of the true relationship, so it cannot represent the signal even in the training data. Unlike overfitting, which looks great in-sample and fails out-of-sample, underfitting looks mediocre everywhere, because the model has not captured the pattern anywhere. It is the high-bias end of the bias-variance trade-off: the error comes from systematic inability to fit the structure, not from sensitivity to noise.

How it differs from overfitting

The two failures are mirror images. Overfitting has low bias and high variance, excellent in-sample fit and poor generalisation. Underfitting has high bias and low variance, poor fit everywhere but stable across samples. The tell-tale distinction is the in-sample result: an overfit model dazzles in-sample, while an underfit model disappoints even there. Diagnosing which one you have determines the fix, because the two require opposite adjustments.

Common causes in strategy design

Underfitting arises from using too few or too weak features, an overly rigid functional form, excessive regularisation, or averaging away the signal with too coarse a timeframe. In the reasonable effort to avoid curve fitting, a researcher may strip a strategy down so far that it can no longer respond to the conditions that actually drive returns. A single moving-average rule applied to an instrument whose edge depends on volatility regime, for instance, is too blunt to capture that dependence.

Why it is easy to miss

Overfitting announces itself with a spectacular backtest that then collapses, which at least draws attention. Underfitting produces a dull, unremarkable backtest that is easy to abandon as simply a bad idea, when in fact a real edge was present but the model was too weak to express it. Because the industry rightly emphasises overfitting, some researchers over-correct and never suspect that their disappointing results are a capacity problem rather than an absence of signal.

Diagnosing underfitting

The diagnostic is that both training and validation performance are poor and close together, indicating high bias rather than high variance. A learning-curve view helps: if adding data does not improve performance and the model already plateaus at a low level on the training set, the ceiling is the model, not the data. Comparing against a slightly more flexible model or a richer feature set, and seeing a genuine, out-of-sample-confirmed improvement, indicates the original was underfitting.

How to fix it without overshooting

Add capacity carefully: introduce better-motivated features, a more flexible but still parsimonious functional form, or relax excessive regularisation. Crucially, add complexity only where an economic rationale suggests real structure, and confirm each addition improves out-of-sample, not just in-sample, performance. The goal is to move from high bias toward the balance point without tipping into high variance. Every step toward more capacity should be validated on untouched data so you do not trade underfitting for overfitting.

Underfitting vs Overfitting

AspectUnderfittingOverfitting
Model complexityToo lowToo high
Bias-varianceHigh bias, low varianceLow bias, high variance
In-sample performancePoorExcellent
Out-of-sample performancePoor and stablePoor and collapsing
Correct fixAdd capacity or better featuresSimplify, regularise, add data

Practical example

Illustrative example (Indian market)

You suspect a volatility-dependent edge in Bank Nifty and, wary of overfitting, you test a single fixed moving-average crossover on capital of Rs 5,00,000. It earns a flat, uninspiring backtest with a Sharpe near 0.3 both in-sample and out-of-sample, and you almost discard the idea. Then you add one economically motivated element, a volatility filter that widens or tightens the rule with the prevailing regime, chosen and validated on separate data. The Sharpe rises to about 0.9 and, importantly, holds out-of-sample. The edge was real all along; the original single-parameter model was simply too blunt to capture it, which is underfitting rather than an absent signal.

Applying one static parameter set to the Nifty across both the low-volatility grind of some years and the sharp, high-volatility episodes of others underfits, because the two regimes behave differently. A model with no way to adapt to India VIX levels averages the regimes together and misses an edge that a modestly more flexible, regime-aware rule can capture.

Limitations

  • Underfitting is easy to confuse with a genuine absence of signal, and the two look similar
  • Fixing it by adding capacity risks overshooting straight into overfitting
  • In finance's low signal-to-noise setting, a weak result may be honest, not underfit
  • Learning-curve diagnostics need enough data to be reliable, which markets often lack
  • Deciding how much complexity a real edge justifies is a judgement, not a formula

Why it matters in practice

  • It is the under-discussed opposite of overfitting and causes real edges to be discarded
  • Over-correcting for overfitting frequently lands researchers in underfitting instead

Common mistakes

  • Stripping a strategy so far to avoid overfitting that it can no longer capture the signal
  • Concluding no edge exists when the model was simply too weak to express it
  • Using excessive regularisation that flattens genuine structure
  • Applying one static parameter set across regimes that behave differently
  • Ignoring that both training and validation results are poor, the signature of high bias
  • Adding capacity without an economic rationale, then over-correcting into overfitting

Professional usage

Skilled researchers diagnose bias and variance separately, reading closely matched but poor training and validation results as underfitting rather than an absent edge. They add capacity deliberately, one economically motivated feature or a modest relaxation of regularisation at a time, and validate each step on untouched data. The discipline is symmetric: just as they distrust a dazzling in-sample result as likely overfit, they distrust a uniformly dull one as possibly a model too blunt to see a real signal.

Key takeaways

  • Underfitting is a model too simple to capture the real signal, so it fails everywhere
  • It is the high-bias mirror image of overfitting and is often missed
  • Its signature is poor performance both in-sample and out-of-sample, close together
  • Over-correcting for overfitting is a common route into underfitting
  • Fix it by adding economically justified capacity and validating each step out-of-sample

Frequently asked questions

What is underfitting in backtesting?
Underfitting is when a strategy or model is too simple or too constrained to capture the genuine structure in the data, so it performs poorly both in-sample and out-of-sample. Unlike overfitting, which dazzles in-sample and fails later, underfitting disappoints everywhere because it never learned the signal.
How is underfitting different from overfitting?
They are mirror images. Overfitting has low bias and high variance, with excellent in-sample fit that collapses out-of-sample. Underfitting has high bias and low variance, with poor fit everywhere but stable across samples. The in-sample result distinguishes them: overfit models look great there, underfit ones do not.
What causes underfitting?
Too few or too weak features, an overly rigid functional form, excessive regularisation, or a timeframe so coarse it averages away the signal. Often it results from over-correcting for overfitting, stripping a strategy down until it can no longer respond to the conditions that drive returns.
Why is underfitting easy to miss?
Because it produces a dull backtest that is easy to dismiss as simply a bad idea, whereas overfitting produces a dramatic result that draws scrutiny. A real edge may be present but hidden by a model too weak to express it, and the disappointing output looks like an absence of signal.
How do I diagnose underfitting?
Look for poor performance that is similar in training and validation, the signature of high bias rather than high variance. If adding data does not help and the model plateaus low even on training data, the limitation is the model. A modestly richer model that improves out-of-sample confirms it.
How do I fix underfitting?
Add capacity carefully: better-motivated features, a more flexible but still parsimonious form, or relaxed regularisation. Add complexity only where an economic rationale suggests real structure, and confirm each addition improves out-of-sample performance so you do not overshoot into overfitting.
Is underfitting worse than overfitting?
Neither is universally worse; they are opposite errors. Overfitting wastes capital on a false edge that collapses live, while underfitting discards a real edge by being too blunt to see it. Both are departures from the bias-variance balance and both cost you, in different ways.
Can avoiding overfitting cause underfitting?
Yes, and it commonly does. The industry rightly warns against overfitting, so some researchers over-simplify and over-regularise until they underfit. Good practice treats both as risks and aims for the balance point, not the simplest possible model regardless of the signal present.
What is the bias-variance trade-off in underfitting?
Underfitting is the high-bias end of the trade-off: error comes from a model too simple to represent the true relationship, not from sensitivity to noise. Reducing underfitting adds capacity to lower bias, accepting a little more variance, until further additions stop improving out-of-sample results.
Does a poor backtest always mean no edge?
No. A poor backtest can mean the idea has no edge, or that the model is too weak to capture an edge that exists. Distinguishing the two requires checking whether a modestly more flexible, economically justified model improves results out-of-sample rather than only in-sample.
How does regime dependence relate to underfitting?
If an edge behaves differently across market regimes, a single static parameter set averages the regimes together and underfits. A model that adapts to regime, for instance to volatility level, can capture the structure the static model misses, provided the added flexibility is validated out-of-sample.
Can machine-learning models underfit?
Yes. Over-regularised models, models with too little capacity, or those given weak features can underfit market data, showing poor and similar training and validation scores. The fix is more capacity or better features, applied carefully and confirmed on held-out data to avoid tipping into overfitting.
How much complexity should I add to fix underfitting?
Only as much as a real edge justifies, and only where an economic rationale points to genuine structure. Add one element at a time and keep it only if it improves out-of-sample performance. The aim is the balance point between bias and variance, not maximum flexibility.
How does underfitting relate to curve fitting?
Curve fitting and overfitting are the same over-tuned failure; underfitting is the opposite under-tuned failure. Where curve fitting fits noise, underfitting misses signal. A sound research process guards against both, moving toward the balance rather than fleeing one error into the other.

Voice search & related questions

Natural-language questions people ask about Underfitting.

What is underfitting in simple terms?
It is using a model too simple to catch the real pattern, like a blunt tool for fine work. It does badly on both old and new data because it never learned the signal in the first place.
How is underfitting different from overfitting?
Overfitting learns too much, including noise, and looks great on old data but fails on new. Underfitting learns too little and looks weak everywhere. They are opposite mistakes.
Why did my simple strategy fail everywhere?
It may be underfitting, meaning it is too simple to capture the edge. If it does poorly on both training and new data, the model, not the idea, might be the problem.
Can being too careful cause underfitting?
Yes. If you strip a strategy down too much to avoid overfitting, you can go too far and miss a real edge. The goal is balance, not the simplest possible model.
How do I fix underfitting?
Add a bit of well-justified complexity, like a better feature or a regime filter, and check it improves results on data you held out. Do it one step at a time so you do not overshoot.
Does a boring backtest mean no edge?
Not necessarily. A dull result can mean the edge is absent, or that your model is too weak to see it. Try a slightly richer model and test it honestly to tell the difference.

Sources & references

    Last reviewed 11 July 2026. Educational content only — not investment advice. Markets and rules change; verify current conventions with SEBI, NSE/BSE and your broker.

    ← Previous
    Overfitting
    Educational content only — not investment advice. Examples use illustrative numbers and simplified models. Backtested results are hypothetical and trading derivatives involves substantial risk. See our Risk Disclosure and SEBI Disclaimer.