Chart 1 from the USD/INR forecasting experiment: 15 models, 1998 to 2051 analysisMOLUSCCAOpen Experiment
Forecasting experiment · USD/INR · 1998 to 2051

Can one year of data
predict twenty-seven?

We trained a model on a single year of the rupee, let it predict every year to 2026 blind, then raced fifteen different models against each other to see which forecasting method actually deserves to be trusted.

Verdict: the model captured the long-run direction with R² of 0.87, missed the timing of every crisis, and was badly overconfident about its own certainty. The simplest model beat all fourteen alternatives, including both machine-learning methods, which finished last. We publish the failures alongside the wins.
Method

What we did, precisely

Three stages, each with the rules fixed before we looked at the answers.

Backtest A. Train on 1999 alone. Predict 2000 to 2026.
Backtest B. Train through 2018. Predict 2019 to 2026.
Horse race. Fifteen models, same blind test.
Forecast. All data. Project 2027 to 2051.

The data gap, stated up front

Our source file of RBI daily reference rates was complete except for 2019, 2020 and 2021. We did not skip them. We predicted them like any other year, then checked against independently referenced published annual averages, marked ref throughout, against file for data present in the source. Years 2022 to 2026 are confirmed directly from the file.

Backtest A

One year in, twenty-seven years out

Trained on 1999 alone: drift 2.57% per year, volatility 1.56% per year. Then released to predict blind.

MAPE
8.4%
average yearly error
MAE
4.3
rupees off
RMSE
5.6
big-miss penalty
R²
0.87
trend captured
Chart 2 from the USD/INR forecasting experiment: 15 models, 1998 to 2051 analysis
Simulated (dashed) against actual (solid). The path runs true through 2013 to 2018 and again 2022 to 2026, but sat far above reality across 2004 to 2012.
Chart 3 from the USD/INR forecasting experiment: 15 models, 1998 to 2051 analysis
Yearly error. Dark bars are misses beyond 10%. The cluster in 2007 to 2011 is the model failing to see the pre-crisis rupee strength and the shocks that followed.
What this actually shows

The direction was right; the journey was wrong. The model assumed steady 2.6% annual depreciation. Across 27 years that average was almost exactly correct, which is why R² is high. But reality was lumpy: the rupee strengthened to 41 in 2007, crashed in 2008, lurched in 2013. A single calm year could not have known those were coming, and no honest model claims otherwise.

YearSimulatedActualError90% bandHit
200044.144.9 file-1.8%[43, 45]out
200145.347.2 file-4.0%[44, 47]out
200246.448.6 file-4.4%[45, 48]out
200347.746.6 file+2.4%[45, 50]in
200448.945.3 file+7.9%[46, 52]out
200550.244.1 file+13.8%[47, 53]out
200651.545.3 file+13.6%[48, 55]out
200752.841.3 file+27.9%[49, 57]out
200854.243.4 file+24.8%[50, 58]out
200955.648.4 file+15.0%[51, 60]out
201057.045.7 file+24.7%[52, 62]out
201158.546.7 file+25.4%[54, 64]out
201260.053.5 file+12.2%[55, 66]out
201361.658.6 file+5.1%[56, 68]in
201463.261.0 file+3.6%[57, 70]in
201564.864.2 file+1.1%[59, 72]in
201666.567.2 file-1.0%[60, 74]in
201768.265.1 file+4.8%[61, 76]in
201870.065.9 file+6.3%[63, 78]in
201971.970.4 ref+2.1%[64, 81]in
202073.774.1 ref-0.5%[66, 83]in
202175.773.9 ref+2.4%[67, 85]in
202277.679.9 file-2.9%[69, 88]in
202379.682.6 file-3.6%[70, 90]in
202481.783.7 file-2.4%[72, 93]in
202583.887.2 file-3.9%[74, 95]in
202685.593.3 file-8.4%[75, 98]in
The overconfidence problem

The number it got dangerously wrong

Its point forecasts were decent. Its confidence was not.

1999 was an unusually calm year, so the model estimated volatility at just 1.56%. True long-run volatility turned out to be 6.40%, more than four times higher. Because it underestimated its own uncertainty, the 90% band was far too narrow and actual values fell outside it roughly half the time.

The lesson

One year is enough to estimate a trend. It is nowhere near enough to estimate risk. A model that is confidently wrong is more dangerous than one that is honestly uncertain. Measuring how much you do not know is harder, and matters more, than the forecast itself.

Backtest B

Trained through 2018, filling the gap

A fairer setup: train on everything to 2018 (drift 2.54%, volatility 6.41%, now honest), predict the missing years and confirm the recent ones.

MAPE
5.6%
average error
MAE
4.8
rupees off
RMSE
5.8
big-miss penalty
Volatility
6.4%
now realistic
Chart 4 from the USD/INR forecasting experiment: 15 models, 1998 to 2051 analysis
Circled points are the referenced gap years the source file was missing. With realistic volatility, the band is finally wide enough to contain reality.
YearSimulatedActualError90% bandHit
201969.970.4 ref-0.7%[65, 75]in
202071.674.1 ref-3.5%[63, 82]in
202173.373.9 ref-0.9%[62, 87]in
202275.079.9 file-6.2%[62, 91]in
202376.882.6 file-7.1%[61, 96]in
202478.583.7 file-6.2%[61, 101]in
202580.387.2 file-7.8%[62, 105]in
202681.793.3 file-12.4%[62, 109]in
The horse race

Fifteen models. The simplest won.

Every model trained on 1998 to 2018 only, then predicted 2019 to 2026 with no peeking. Ranked purely on out-of-sample error, never on how well they fit the past.

Chart 5 from the USD/INR forecasting experiment: 15 models, 1998 to 2051 analysis
Blind-test accuracy. Black is the winner, mid-greys the strong simple models, light grey the failures, including both machine-learning methods.
RankModelMAPERMSEMAE
1GBM (drift) view8.53%7.807.10
2Quadratic view8.67%7.897.18
3Exponential view10.29%9.098.50
4Logistic (S-curve) view10.29%9.098.50
5Ornstein-Uhlenbeck view10.86%10.039.06
6Random walk + drift view11.38%10.529.49
7ARIMA monthly->yearly view13.48%13.4611.41
8ARIMA(2,1,2) view14.13%13.4211.86
9Holt (exp smoothing) view14.33%13.2811.97
10Holt (damped) view14.71%13.6412.29
11Linear view15.22%13.3512.57
12ARIMA(1,1,1) view16.49%15.5813.82
13Naive (last value) view17.67%16.4014.76
14Gradient Boosting view17.68%16.4014.77
15Random Forest view17.84%16.5014.89

Click any model for its full explanation: how it works, where it fits, and where it fails.

No model fits every scenario

This is the point most easily missed. Every model on that list is the correct choice for some problem and the wrong choice for others. Random Forest finished last here and would comfortably win on a large tabular fraud dataset. The logistic S-curve failed on a currency and is the right tool for technology adoption. Ornstein-Uhlenbeck was wrong for a trending rupee and is standard for interest rates. Machine learning did not lose because it is bad; it lost because 21 training points and a target moving beyond every value it had seen is the precise condition under which tree models cannot function.

Choosing a model is choosing a shape. The job is matching the shape of the method to the shape of the problem, not reaching for whatever is most sophisticated.

Forecast

2027 to 2051, and how much to trust it

Refit on the full history: drift 3.53% per year, volatility 6.40%, starting from 96.2.

Chart 6 from the USD/INR forecasting experiment: 15 models, 1998 to 2051 analysis
Full history and the median forecast with 90% band. The band widens enormously with time, and that width is the model being honest, not vague.
YearMedian90% bandWidth
202798[91, 106]15
2030108[89, 132]43
2035128[94, 174]80
2040151[103, 223]120
2045179[113, 281]168
2051219[129, 366]237
How much confidence

The central line is a reasonable extrapolation of a stable 25-year trend, so roughly 219 by 2051 is defensible as a central estimate.

But the honest reading is the band, not the line. By 2051 the 90% range spans 129 to 366. A quarter century of unknown crises and policy shifts lives inside that width. Anyone quoting a single confident 2051 number is not being straight with you.

Verdict

Did our simulation work?

Yes for trend, no for timing, and it taught us more by failing than it would have by quietly succeeding.

What worked

Long-run direction captured with R² 0.87. Beat the naive benchmark by more than half. Survived a blind race against fourteen alternatives and won it.

What failed

Missed the timing of every major shock. Starved of data it was badly overconfident, with real values outside its 90% band half the time.