Sam Austinon November 2, 202613 min read

Direct vs Recursive Multi-Step Forecasting: Which Strategy Wins?

Direct vs Recursive Multi-Step Forecasting: Which Strategy Wins?
Contents

Figure 1: Direct vs Recursive Multi-Step Forecasting: Which Strategy Wins? — one model applied 24 times, or 24 models applied once each

Everyone wants a winner. I built both strategies by hand and ran them on three synthetic series. The honest result: on two of them they tied, and on the third the winner depended on the horizon.

The Two Strategies in One Minute

Recursive: train one model to predict the next step. To forecast 24 steps, feed each prediction back in as the newest lag and repeat. One model, applied 24 times.

Direct: train a separate model for each horizon. Model 7 learns to predict seven steps ahead straight from the last known values. 24 steps means 24 models.

The review by Ben Taieb, Bontempi, Atiya and Sorjamaa describes the same split: recursive "iterates, H times, a one-step ahead forecasting model," while direct forecasts each horizon "independently from the others" [1]. The paper also covers MIMO (one model that returns the whole future vector) and DirRec (direct models that also receive earlier forecasts).

How I Tested Them

I used a 24-step horizon and 24 lags, plus sine and cosine of the time of day, which are known for future steps. I scored rolling forecasts on the last 25% of each series, using MAE per horizon, over 227 forecast origins.

  • Series A: linear, with strong autoregression and a 24-step seasonal cycle.
  • Series B: a threshold model, where the rule changes with the sign of the last value.
  • Series C: the logistic map, a chaotic series that is almost deterministic one step ahead and unpredictable far ahead. This one is a stress test, not a realistic dataset.

Models: Ridge regression and gradient boosting, each in both strategies.

Result 1: On Well-Behaved Series, It Is a Tie

Series / model Recursive mean MAE Direct mean MAE
A, Ridge 1.254 1.283
A, HGB 1.459 1.538
B, Ridge 0.250 0.251
B, HGB 0.270 0.272

The gaps run from 0.4% on series B to about 5% on series A with boosting, and recursive was ahead in all four cases. That is not a decisive win. It is the kind of difference a different seed could reverse.

Per horizon the picture is just as flat. On series A, boosting was better recursively at 6 steps, 1.469 versus 1.618, and still better at 24, 1.439 versus 1.509. On series B the sign flipped only at the final step, 0.270 versus 0.264. With one run and 227 origins, I would not read much into those swings.

Why No Error Pile-Up?

On series A and B, errors barely grew with the horizon at all. Series B's recursive Ridge error was 0.210 at one step and 0.249 at 24 steps.

These series pull back toward their mean, so a recursive forecast has little room to drift. Forecast uncertainty still grows with the horizon in general, which is why prediction intervals widen as you go further out [3], but a strongly mean-reverting series hides that growth in the point forecast.**

Result 2: Where Recursion Does Compound

Series C shows what people warn about, and it shows it later than I expected:

Horizon Recursive MAE Direct MAE Mean predictor
1 0.002 0.002 0.171
6 0.012 0.013 0.169
12 0.101 0.167 0.172
17 0.206 0.199 0.170
24 0.238 0.194 0.172

Three things stand out.

  1. Short horizons: the two are identical. 0.002 at one step, 0.013 at six.
  2. The middle belongs to recursion. At 12 steps recursive scored 0.101 against 0.167 for direct.
  3. The crossover arrives at step 17. From there direct wins, and at 24 steps it is 0.194 against 0.238.

Both end up worse than simply predicting the series mean, which scores 0.172. At the far end of a chaotic horizon, no model is better than the average.

Why does recursion hold on so long? The one-step map is nearly perfectly learnable, 0.002 at step 1, so iterating it reproduces the true dynamics for about sixteen steps. A direct model at step 12 is asked to learn a 12-step map that chaos has already destroyed, so it falls back toward the mean long before recursion does. Direct does not rescue you; it just degrades more gracefully at the end.

I would not generalise from a chaotic map to your sales data. It shows the mechanism, not how common it is.

Result 3: The Cost Difference Is Real

In my runs, gradient boosting trained 1 model in about 0.2 seconds for recursive and 24 models in about 6.4 seconds for direct, a factor of roughly 27.

That ratio makes sense, since direct needs one fit per horizon, and the cost grows with the horizon. A 96-step horizon at 15-minute resolution would mean 96 fits.

Recursive pays at prediction time instead: it must run steps in sequence. Direct predictions can run in parallel.

What About MIMO?

MIMO trains one model that outputs the whole forecast vector. I checked one detail: with Ridge on identical features, a multi-output fit gave the same predictions as 24 separate Ridge fits, with a maximum difference of 3.2e-13 [4]. Scikit-learn's MultiOutputRegressor describes the same idea, "fitting one regressor per target" [2]. So for Ridge, MIMO and direct are the same thing. Tree models that predict vectors natively may behave differently, but I did not test any.

The review's conclusion is worth knowing: "The Multiple-Output strategies (MIMO and DIRMO) are invariably the best strategies." Among the single-output strategies, recursive "has almost always a smaller SMAPE and a better ranking than the DIR strategy," and DirRec "is the worse strategy overall" [1]. The authors used 111 daily ATM series and one base learner, Lazy Learning, and the paper itself notes that earlier comparisons contradicted each other. Treat it as one strong data point, not a verdict.

How I Would Choose

  1. Start with recursive for short horizons or stable series. It is cheap, and in my tests it matched direct on well-behaved data.
  2. Try direct when the horizon is long, the dynamics are nonlinear, or you can afford one model per step.
  3. Test MIMO or DirRec if you have a model that handles vector outputs well and your steps depend on each other.
  4. Always compare per-horizon error, not one average. The average favoured recursive on series C, while the last eight horizons favoured direct.
  5. Check the cost of H models against the gain.
  6. Use the same rolling origins for every strategy, so you compare like with like.

The best habit is to treat the strategy as a hyperparameter. You would not guess a tree depth, so do not guess this either.

Common Pitfalls

  1. Reporting only the mean MAE. It hid an eight-horizon stretch where direct was better.
  2. Assuming recursion always compounds. On mean-reverting series it barely moved, 0.210 to 0.249.
  3. Assuming direct always wins at long horizons. It lost badly at 12 steps on the chaotic series.
  4. Comparing strategies on different origins or different features. Then you are measuring the setup, not the strategy.
  5. Forgetting that direct needs one target column per horizon. Row counts shrink as the horizon grows, and the last rows of each target vanish.
  6. Ignoring fit time when the horizon is large. 24 fits is fine, 96 is not free.

Best Book to Read

CoverBookDescriptionGet it
Cover of “Practical Time Series Analysis” Practical Time Series Analysisby Aileen Nielsen covers multi-step forecasting strategies with runnable Python, including the recursive and direct trade-off this post measures, and treats model choice as something you test rather than assume. View on Amazon

As an Amazon Associate I earn from qualifying purchases. Links on this page are affiliate links, and I may earn a commission at no extra cost to you.

Wrapping Up

Which strategy wins? In my experiments, mostly neither: the means tied on two series, and on the chaotic one recursive held the middle of the horizon while direct took the end. What differed reliably was cost, 1 model versus 24.

Take one forecast you run today, implement both strategies in under 50 lines, and plot MAE by horizon. If the lines cross, you have learned where each strategy belongs in your pipeline.

Frequently Asked Questions

What is the difference between recursive and direct multi-step forecasting?

Recursive trains one model to predict the next step and feeds each prediction back in, so a 24-step forecast is one model applied 24 times. Direct trains 24 separate models, one per horizon, each predicting its own step straight from the last known values. The review by Ben Taieb and colleagues describes the same split.

Why does recursive forecasting compound errors?

Because every step consumes the previous prediction instead of real data. A small error at step one becomes an input at step two, and the model has never seen that input during training. In my logistic-map test the damage only showed after about 16 steps, because the learned one-step map is accurate enough to keep the trajectory on track until then.

When should I use direct forecasting instead?

When the horizon is long, the dynamics are nonlinear, or you can afford one fit per step. In my tests direct was not automatically better in the middle of the horizon: on the chaotic series it lost badly at 12 steps and only took over at step 17. Compare per-horizon error before deciding.

Is MIMO the same as direct forecasting?

For Ridge it is. A multi-output Ridge fit gave the same predictions as 24 separate Ridge fits, with a maximum difference of 3.2e-13. Scikit-learn's MultiOutputRegressor describes the same construction, fitting one regressor per target, so for linear models the two are the same thing.

How do I compare recursive and direct fairly?

Use the same rolling origins for both, the same features, and the same training window. Then score MAE per horizon rather than averaging, because the average hides the crossover. In my chaotic test the mean favoured recursive while the last eight horizons favoured direct.

Sources

What are You Looking For?

esc